Datasets:

bskang
/

CVPR2023_title_abstract_intro

Dataset card Viewer Files Files and versions Community

Split (1)

train · 2.34k rows

title stringlengths 28 135	abstract stringlengths 0 12k	introduction stringlengths 0 12k
Weers_Masked_Autoencoding_Does_Not_Help_Natural_Language_Supervision_at_Scale_CVPR_2023	Abstract Self supervision and natural language supervision have emerged as two exciting ways to train general purpose im- age encoders which excel at a variety of downstream tasks. Recent works such as M3AE [ 31] and SLIP [ 63] have sug- gested that these approaches can be effectively combined, but most notably their results use small ( <20M examples) pre-training datasets and don’t effectively reﬂect the large- scale regime ( >100M samples) that is commonly used for these approaches. Here we investigate whether a similar approach can be effective when trained with a much larger amount of data. We ﬁnd that a combination of two state of the art approaches: masked auto-encoders, MAE [ 37] and contrastive language image pre-training, CLIP [ 68] provides a beneﬁt over CLIP when trained on a corpus of 11.3M image-text pairs, but little to no beneﬁt (as evaluated on a suite of common vision tasks) over CLIP when trained on a large corpus of 1.4B images. Our work provides some much needed clarity into the effectiveness (or lack thereof) of self supervision for large-scale image-text training.	1. Introduction Large scale pretraining has become a powerful tool in the arsenal of computer vision researchers to produce state of the art results across a wider variety of tasks [ 39,88,95,98]. However, when pre-training on tens of millions to billions of images it is difﬁcult to rely on standard supervised meth- ods to train models, as datasets of this size often lack re- liable labels. In the presence of these massive but largely under-curated datasets, two general classes of methods to train general purpose image encoders have emerged: 1.Self Supervised techniques that learn visual represen- tations from the image data alone [ 11,36] 2.Natural Language Supervised methods that utilize paired free-form text data to learn visual representa- tions [ 43,69] Due to the unique strengths and weaknesses of each ap-proach1, a recent ﬂurry of work has introduced methods that combine both forms of supervision [ 31,56,64,78] to vary- ing degrees of success. While each of these methods estab- lishes some regime where the additional supervision helps, none of these “joint-supervision” methods advance state of the art in any meaningful way. Additionally, to our knowl- edge none of these methods have shown comparative results at the scale many large scale vision models are currently trained at ( >100M examples) [ 43,66,69,73,80,82,98]. Fur- thermore, methods that use both forms of supervision start with the presumption that the additional supervision is help- fuland either often lack clean ablations or lack evaluations in a “high accuracy” regime—leading to further confusion regarding whether a combination of these methods can ac- tually improve the state of the art. To clarify this issue, in this work, we investigate a simple question: Does a combination of self supervision and natu- ral language supervision actually lead to higher quality visual representations? In order to answer this, we ﬁrst introduce a straight- forward baseline approach that combines standard self su- pervision and language supervision techniques. We com- bine masked auto-encoders (MAE) and contrastive lan- guage image-pretraining (CLIP) to make MAE-CLIP. We then present a careful study of the performance of MAE, M3AE, CLIP and MAE-CLIP across a wide variety of tasks in two distinct regimes: a “low-sample”211.3 million example regime and a “high-sample” 1.4 billion example regime. We train self-supervised and language-supervised methods using the same pre-training datasets under the as- sumption that we have no knowledge about downstream tasks. Our experiments show: 1.In the low sample size regime, without changing the ﬁnal pooling operation in the network, we observe a large performance improvement, namely 6% on Ima- geNet [ 18] and 4% on VTAB [ 105]. However, when 1Self supervised methods can learn representations without labels, but natural language supervision learns better representations. Natural lan- guage supervised methods rely on quality of captions 2We note that what low sample means has changed substantially over the last few years 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23432 we modify the pooling operation, the improvement substantially decreases to around 1% on both Ima- geNet and VTAB. 2.In the high sample size regime, there is virtually no dif- ference in performance between MAE-CLIP and CLIP across ImageNet, VTAB, and VQA tasks. We believe our work is the ﬁrst careful study of this form and contextualizes recent progress in both self-supervision and natural language supervision. The rest of the paper is organized as follows: In Sec- tion2, we cover related work in the areas of self supervision and natural language supervision. In Section 3, we give an overview of the baseline methods we study, MAE, M3AE, CLIP and our new baseline MAE-CLIP. Then we present and analyse our small scale and large scale experimental ﬁndings in Sections 4and5. Finally, we discuss potential explanations for our ﬁndings and some future work in 6.
Wang_SmartAssign_Learning_a_Smart_Knowledge_Assignment_Strategy_for_Deraining_and_CVPR_2023	Abstract Existing methods mainly handle single weather types. However, the connections of different weather conditions at deep representation level are usually ignored. These con- nections, if used properly, can generate complementary rep- resentations for each other to make up insufficient train- ing data, obtaining positive performance gains and better generalization. In this paper, we focus on the very corre- lated rain and snow to explore their connections at deep representation level. Because sub-optimal connections may cause negative effect, another issue is that if rain and snow are handled in a multi-task learning way, how to find an optimal connection strategy to simultaneously improve de- raining and desnowing performance. To build desired con- nection, we propose a smart knowledge assignment strat- egy, called SmartAssign, to optimally assign the knowledge learned from both tasks to a specific one. In order to fur- ther enhance the accuracy of knowledge assignment, we propose a novel knowledge contrast mechanism, so that the knowledge assigned to different tasks preserves better uniqueness. The inherited inductive biases usually limit the modelling ability of CNNs, we introduce a novel trans- former block to constitute the backbone of our network to ef- fectively combine long-range context dependency and local image details. Extensive experiments on seven benchmark datasets verify that proposed SmartAssign explores effec- tive connection between rain and snow, and improves the performances of both deraining and desnowing apparently. The implementation code will be available at https:// gitee.com/mindspore/models/tree/master/ research/cv/SmartAssign .	1. Introduction Bad weather types, such as haze, rain, and snow in- evitably degrade the visual quality of images, meanwhile decrease the performances of other downstream computer Input Restormer [58] Ours Input HDCWNet [7] Ours Figure 1. Given challenging rainy (with blurry rain streaks) and snowy (with high bright snowflakes) images , the proposed method effectively removes the artifacts of rain and snow simultaneously, achieving better results than the state-of-the-art approaches. This is attributed to the unique knowledge which captures accurate fea- tures of rain/snow as well as the common knowledge boosting the generalization of our model to real data. vision tasks, e.g., autonomous driving [57]. Existing meth- ods mainly focus on single weather types, e.g., deraining [14,19,26,45,46,48–50,56,60], dehazing [5,10,31,41,54], and desnowing [6, 7, 32, 49]. However, these methods usu- ally ignore the connections among these weather types, which, if used properly, may simultaneously improve the performance of multiple image recovery tasks. Some methods attempt to explore the connections among different weather types by handling them with an uni- fied architecture and one set of pre-trained weights, e.g., [8, 25, 27]. But they neglect the difference of multiple weather types, the uniqueness belonging to single weather types may harm the performance of other weather recovery tasks. Therefore, the performances of such unified networks This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3677 are usually lower than the ones for single weather types [8]. In this paper, we focus on the very similar rain and snow and develop a novel Multi-Task Learning (MTL) strategy to explore their connections at deep representation level, meanwhile avoiding the uniqueness of one weather type from damaging the performance of the other weather re- covery task. Specifically, our goal is to accurately find the representations (i.e., connections) which can be shared by deraining and desnowing to simultaneously enhance both their performance, and meanwhile determine the exclusive representations (i.e., uniqueness) for one task to specially promote its own performance as well as avoid such unique- ness from damaging the other task. To facilitate the descrip- tion of our method, we define the deep representation of a single network channel as a knowledge atom . All the knowl- edge atoms constitute the knowledge learned by networks. Similar to conventional MTL method, we also use a backbone encoder E(·)to learn the whole image-recovery knowledge simultaneously from rainy and snowy images, and two task-targeted decoders Ddr(·)andDds(·)are fol- lowed in parallel to remove rain and snow, separately. Con- ventional MTL takes the whole knowledge as the input for the subsequent decoders. Though such mechanism makes the best of the connections between both tasks, the in- fluences of the uniqueness of single tasks are neglected, i.e., the uniqueness of rain may harm the performance of desnowing, and vice versa. Instead, we propose a novel Gated Knowledge Filtering Module (GKFM) to select op- timal knowledge atoms for both tasks via a highly smart strategy, so that the connections between both tasks are suf- ficiently explored and the uniqueness of single tasks is prop- erly used. To coordinate with GKFM, we design a Task- targeted Knowledge FeedForward mechanism (TKFF) to let every knowledge atom flow to its related tasks. Through our GKFM and TKFF, we realize a smart knowledge assign- ment, in which both tasks adaptively explore their connec- tions and uniqueness via gradient backward-propagation. Hence, we term our MTL mechanism as SmartAssign . In order to further enhance the accuracy of knowledge assignment, i.e., toward optimally exploring the connec- tions and uniqueness of deraining and desnowing, we in- troduce a novel knowledge contrast, making the same kind of knowledge atoms more closer, and the ones belonging to different kinds more discriminative under a similarity met- ric. In this process, all the knowledge atoms are transformed into a new low-dimension feature space by a Dimension Re- duction Module (DRM) to avoid model collapse when op- erating on the original high-dimension knowledge atoms. Currently, CNNs are still the mainstream choice for image recovery. However, the inherited inductive biases limit their modelling capacity for long-range context depen- dency. Though they can also obtain a large receptive field by stacking a deep architecture, such indirect modelling isindeed inferior to that of a transformer, which models both short and long range dependency directly via self-attention. In this paper, we adopt transformer blocks to constitute our backbone encoder E(·). Usually, transformer needs suffi- cient training pairs to ensure good performance. Hence, our transformer block introduces a gated CNN branch to complement limited training data via the inductive biases. Moreover, the locality of CNN helps to recover degraded image details and the gated operation is used to reduce re- dundant features caused by the combination of CNN and transformer. Figure 1 gives two examples of deraining and desnowing. By contrast, our method obtains better image recovery quality on both tasks than SOTA methods. Our contributions are summarized in the following: • We propose a novel knowledge assignment strategy, i.e., SmartAssign, to excavate the connections and uniqueness of rain and snow, so that their connections are used to enhance the performance of both tasks and the uniqueness is applied to boost corresponding task and avoided from damaging the other task. • We propose a novel knowledge contrast mechanism to further boost the accuracy of knowledge assignment, in which a dimension reduction module (DRM) is in- troduced to stabilize the training of our model. • We propose a novel transformer block to make the best use of the superiority of self-attention and convolution, in which gated operations are introduced to alleviate the feature redundancy.
Wang_PDPPProjected_Diffusion_for_Procedure_Planning_in_Instructional_Videos_CVPR_2023	Abstract In this paper, we study the problem of procedure plan- ning in instructional videos, which aims to make goal- directed plans given the current visual observations in un- structured real-life videos. Previous works cast this prob- lem as a sequence planning problem and leverage either heavy intermediate visual observations or natural language instructions as supervision, resulting in complex learning schemes and expensive annotation costs. In contrast, we treat this problem as a distribution fitting problem. In this sense, we model the whole intermediate action sequence distribution with a diffusion model (PDPP), and thus trans- form the planning problem to a sampling process from this distribution. In addition, we remove the expensive inter- mediate supervision, and simply use task labels from in- structional videos as supervision instead. Our model is a U-Net based diffusion model, which directly samples ac- tion sequences from the learned distribution with the given start and end observations. Furthermore, we apply an ef- ficient projection method to provide accurate conditional guides for our model during the learning and sampling pro- cess. Experiments on three datasets with different scales show that our PDPP model can achieve the state-of-the- art performance on multiple metrics, even without the task supervision. Code and trained models are available at https://github.com/MCG-NJU/PDPP.	1. Introduction Instructional videos [1,31,38] are strong knowledge car- riers, which contain rich scene changes and various actions. People watching these videos can learn new skills by fig- uring out what actions should be performed to achieve the desired goals. Although this seems to be natural for hu- mans, it is quite challenging for AI agents. Training a model that can learn how to make action plans to transform from the start state to goal is crucial for the next-generation AI system as such a model can analyze complex human be- B: Corresponding author (lmwang@nju.edu.cn). (b) Supervised by visual observations (c) Supervised by task class (Ours)Task : Make Meringue Seen observations Intermediate supervision Predicted action Pour eggAdd SugarWhisk mixture (a) Supervised by language instructions Whisk mixturePour eggAdd SugarSpread mixtureWhisk mixturePour eggAdd SugarSpread mixture Whisk mixturePour eggAdd SugarSpread mixtureFigure 1. Procedure planning example. Given a start observation ostart and a goal state ogoal, the model is required to generate a sequence of actions that can transform ostart toogoal. Previous approaches rely on heavy intermediate supervision during training, while our model only needs the task class labels (bottom row). haviours and help people with goal-directed problems like cooking or repairing items. Nowadays the computer vision community is paying growing attention to the instructional video understanding [4, 8, 9, 24, 37]. Among them, Chang et al. [4] proposed a problem named as procedure planning in instructional videos, which requires a model to produce goal-directed action plans given the current visual observa- tion of the world. Different with traditional procedure plan- ning problem in structured environments [12, 29], this task deals with unstructured environments and thus forces the model to learn structured and plannable representations in real-life videos. We follow this work and tackle the proce- dure planning problem in instructional videos. Specifically, given the visual observations at start and end time, we need to produce a sequence of actions which transform the envi- ronment from start state to the goal state, as shown in Fig. 1. Previous approaches for procedure planning in instruc- tional videos often treat it as a sequence planning prob- lem and focus on predicting each action accurately. Most works rely on a two-branch autoregressive method to pre- dict the intermediate states and actions step by step [2,4,30]. Such models are complex and easy to accumulate errors during the planning process, especially for long sequences. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14836 Recently, Zhao et al . [36] proposed a single branch non- autoregressive model based on transformer [33] to predict all intermediate steps in parallel. To obtain a good per- formance, they used a learnable memory bank in the trans- former decoder, augmented their model with an extra gen- erative adversarial framework [13] and applied a Viterbi post-processing method [34]. This method brought multiple learning objectives, complex training schemes and tedious inference process. Instead, we assume procedure planning as a distribution fitting problem and planning is solved with a sampling process. We aim to directly model the joint dis- tribution of the whole action sequence in instructional video rather than every discrete action. In this perspective, we can use a simple MSE loss to optimize our generative model and generate action sequence plans in one shot with a sam- pling process, which results in less learning objectives and simpler training schemes. For supervision in training, in addition to the action se- quence, previous methods often require heavy intermediate visual [2,4,30] or language [36] annotations for their learn- ing process. In contrast, we only use task labels from in- structional videos as a condition for our learning (as shown in Fig. 1), which could be easily obtained from the key- words or captions of videos and requires much less labeling cost. Another reason is that task information is closely re- lated to the action sequences in a video. For example, in a video of jacking up a car , the possibility for action add sugar appears in this process is nearly zero. Modeling the uncertainty in procedure planning is also an important factor that we need to consider. That is, there might be more than one reasonable plan sequences to trans- form from the given start state to goal state. For example, change the order of add sugar andadd butter inmaking cake process will not affect the final result. So action se- quences can vary even with the same start and goal states. To address this problem, we consider adding randomness to our distribution-fitting process and perform training with a diffusion model [18, 26]. Solving procedure planning prob- lem with a diffusion model has two main benefits. First, a diffusion model changes the goal distribution to a random Gaussian noise by adding noise slowly to the initial data and learns the sampling process at inference time as an iter- ative denoising procedure starting from a random Gaussian noise. So randomness is involved both for training and sam- pling in a diffusion model, which is helpful to model the un- certain action sequences for procedure planning. Second, it is convenient to apply conditional diffusion process with the given start and goal observations based on diffusion models, so we can model the procedure planning problem as a con- ditional sampling process with a simple training scheme. In this work, we concatenate conditions and action sequences together and propose a projected diffusion model to perform conditional diffusion process.Contributions . To sum up, the main contributions of this work are as follows: a) We cast the procedure plan- ning as a conditional distribution-fitting problem and model the joint distribution of the whole intermediate action se- quence as our learning objective, which can be learned with a simple training scheme. b) We introduce an efficient ap- proach for training the procedure planner, which removes the supervision of visual or language features and relies on task supervision instead. c) We propose a novel projected diffusion model (PDPP) to learn the distribution of action sequences and produce all intermediate steps at one shot. We evaluate our PDPP on three instructional videos datasets and achieve the state-of-the-art performance across differ- ent prediction time horizons. Note that our model can still achieve excellent results even if we remove the task super- vision and use the action labels only.
Wu_High-Fidelity_3D_Face_Generation_From_Natural_Language_Descriptions_CVPR_2023	Abstract Synthesizing high-quality 3D face models from natural language descriptions is very valuable for many applica- tions, including avatar creation, virtual reality, and telep- resence. However, little research ever tapped into this task. We argue the major obstacle lies in 1) the lack of high- quality 3D face data with descriptive text annotation, and 2) the complex mapping relationship between descriptive language space and shape/appearance space. To solve these problems, we build DESCRIBE 3Ddataset, the first large-scale dataset with fine-grained text descriptions for text-to-3D face generation task. Then we propose a two- stage framework to first generate a 3D face that matches the concrete descriptions, then optimize the parameters in the 3D shape and texture space with abstract description to refine the 3D face model. Extensive experimental re- sults show that our method can produce a faithful 3D face that conforms to the input descriptions with higher accu- racy and quality than previous methods. The code and DE- SCRIBE 3Ddataset are released at https://github. com/zhuhao-nju/describe3d .	1. Introduction 3D faces are highly required in many cutting-edge tech- nologies like digital humans, telepresence, and movie spe- cial effects, while creating a high-fidelity 3D face is very complex and requires vast time from an experienced mod- eler. Recently, many efforts are devoted to the synthesis of text-to-image and image-to-3D, but they lack the ability to synthesize 3D faces given an abstract description. However, there is still no reliable solution to synthesize high-quality 3D faces from descriptive texts in natural language. We consider the difficulties of synthesizing high-quality 3D face models from natural language descriptions lie in two folds. Firstly, there is still no available fine-grained dataset that contains 3D face models and corresponding text descriptions in the research community, which is crucial for training learning-based 3D generators. Beyond that, it is difficult to leverage massive 2D Internet images to learn Figure 1. Given a text describing the appearance ( left), our method can synthesize high-quality 3D faces ( middle ) containing 3D mesh and textures. The resulting model can be easily processed into a rigged face with hair and accessories ( right ). The dark blue texts indicate concrete descriptions and the brown texts indicate abstract descriptions, and similarly hereinafter. high-quality text-to-3D mapping. Secondly, cross-modal mapping from texts to 3D models is non-trivial. Though the progress made in text-to-image synthesis is instructive, the problem of mapping texts to 3D faces is even more chal- lenging due to the complexity of 3D representation. In this work, we aim at tackling the task of high-fidelity 3D face generation from natural text descriptions from the above two perspectives. We first build a 3D-face-text dataset (named D ESCRIBE 3D), which contains 1,627high- quality 3D faces from HeadSpace dataset [6] and FaceScape dataset [48, 55], and fine-grained manually-labeled facial features. The provided annotations include 25 facial at- tributes, each of which contains 3 to 8 options describing the facial feature. Our dataset covers various races and ages and is delicate in 3D shape and texture. We then propose a two-stage synthesis pipeline, which consists of a concrete synthesis stage mapping the text space to the 3D shape and texture space, and an abstract synthesis stage refining the 3D face with a prompt learning strategy. The mapping for different facial features is disentangled and the diversity of the generative model can be controlled by the additional input of random seeds. As shown in Figure 1, our pro- posed model can take any word description or combination This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4521 of phrases as input, and then generate an output of a fine- textured 3D face with appearances matching the descrip- tion. Extensive experiments further validate that the con- crete synthesis can generate a detailed 3D face that matches the fine-grained descriptive texts well, and the abstract syn- thesis enables the network to synthesize abstract features like “wearing makeup” or “looks like Tony Stark”. In summary, our contributions are as follows: • We explore a new topic of constructing a high-quality 3D face model from natural descriptive texts and pro- pose a baseline method to achieve such a goal. • A new dataset - D ESCRIBE 3D is established with de- tailed 3D faces and corresponding fine-grained de- scriptive annotations. The dataset will be released to the public for research purposes. • The reliable mapping from the text embedding space to the 3D face parametric space is learned by intro- ducing the descriptive code space as an intermediary, which forms the core of our concrete synthesis mod- ule. Region-specific triplet loss and weighted ℓ1loss further boost the performance. • Abstract learning based on CLIP is introduced to fur- ther optimize the parametric 3D face, enabling our re- sults to conform with abstract descriptions.
Xie_Blemish-Aware_and_Progressive_Face_Retouching_With_Limited_Paired_Data_CVPR_2023	Abstract Face retouching aims to remove facial blemishes, while at the same time maintaining the textual details of a giv- en input image. The main challenge lies in distinguishing blemishes from the facial characteristics, such as moles. Training an image-to-image translation network with pixel- wise supervision suffers from the problem of expensive paired training data, since professional retouching need- s specialized experience and is time-consuming. In this paper, we propose a Blemish-aware and Progressive Face Retouching model, which is referred to as BPFRe. Our framework can be partitioned into two manageable stages to perform progressive blemish removal. Speciﬁcally, an encoder-decoder-based module learns to coarsely remove the blemishes at the ﬁrst stage, and the resulting interme- diate features are injected into a generator to enrich lo- cal detail at the second stage. We ﬁnd that explicitly sup- pressing the blemishes can contribute to an effective col- laboration among the components. Toward this end, we incorporate an attention module, which learns to infer a blemish-aware map and further determine the correspond- ing weights, which are then used to reﬁne the intermediate features transferred from the encoder to the decoder, and from the decoder to the generator. Therefore, BPFRe is able to deliver signiﬁcant performance gains on a wide range of face retouching tasks. It is worth noting that we reduce the dependence of BPFRe on paired training samples by impos- ing effective regularization on unpaired ones.	1. Introduction With the development of social media, there is an in- creased demand for facial image beautiﬁcation from selﬁes to portraits and beyond. Facial skin retouching aims to re- Corresponding author. Figure 1. Visual comparison of the activation maps produced by a generic attention module [47] ( second row ) and the blemish-aware module used in BPFRe ( third row ), given a number of images of faces with blemishes ( top row ). BPFRe is capable of applying attention on the regions close to the manual retouching regions (bottom row ). move any unexpected blemishes from facial images, while preserving the stable characteristics that associate with face identity [2, 38, 41]. The main challenge is due to the wide range of blemishes including from small spots to severe ac- ne. Conventional methods are based on blind smoothing, such that the facial characteristics, such as moles and freck- les, may be removed. Professional face retouching can be expensive and needs specialized experience, which impedes the collection of large-scale paired data for model training. Deep neural networks have been widely used for image- to-image translation, especially based on Generative Adver- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5599 sarial Networks (GANs) [6, 12, 22]. The translation per- formance has witnessed rapid progress in style transfer [7], image restoration [45, 48], image inpainting [47], and so on. The existing models are typically based on an encoder- decoder architecture. The source image is encoded into a latent representation, based on which a task-speciﬁc trans- formation is performed by the decoder. Different from the above image enhancement tasks, the regions needed to be retouched may be small, and most of the pixels are un- changed in this case. Generic encoder-decoder-based trans- lation methods can preserve irrelevant content but tend to overlook large blemishes and produce over-smoothed im- ages. Considering that StyleGAN-based methods have the capability of rendering the complex textual details [1, 11], we design a two-stage progressive face retouching frame- work to make use of the advantage of these types of archi- tectures, and learn the blemish-aware attention (as shown in Figure 1) to guide the image rendering process. More speciﬁcally, we propose a Blemish-aware Progres- sive Face Retouching model (BPFRe), which consists of t- wo stages: An encoder-decoder architecture is applied at the ﬁrst stage to perform coarse retouching. The intermediate features from the encoder are integrated into the decoder via skip connections for better reconstruction of image content. At the second stage, we modify the generator architecture of StyleGAN [22] to operate on the multi-scale intermedi- ate features of the decoder and render an image with ﬁner details. We consider that blemish removal cannot be ef- fectively achieved by simply transferring the intermediate features between the components, since there is no mecha- nism to suppress the blemishes before being passed to the next components. To address this issue, we incorporate t- wo blemish-aware attention modules between the encoder and decoder, and between the decoder and generator, re- spectively. This design enables progressive retouching by leveraging and reﬁning the information from the previous components. In addition to the paired training images, we use the unpaired ones to optimize the discriminator, which in turn guides the generator to synthesize realistic detail- s. We perform extensive experiments to qualitatively and quantitatively assess BPFRe on both standard benchmarks and data in the wild. The main contributions of this work are summarized as follows: (a) To deal with a wide range of facial blemishes, we exploit the merits of both encoder-decoder and genera- tor architectures by seamlessly integrating them into a uni- ﬁed framework to progressively remove blemishes. (b) A blemish-aware attention module is incorporated to enhance the collaboration between the components by reﬁning the intermediate features that are transferred among the com- ponents. (c) We leverage unpaired training data to regular- ize the proposed framework, which effectively reduces the dependence on paired training data.
Walz_Gated_Stereo_Joint_Depth_Estimation_From_Gated_and_Wide-Baseline_Active_CVPR_2023	Abstract We propose Gated Stereo, a high-resolution and long- range depth estimation technique that operates on active gated stereo images. Using active and high dynamic range passive captures, Gated Stereo exploits multi-view cues alongside time-of-flight intensity cues from active gating. To this end, we propose a depth estimation method with a monocular and stereo depth prediction branch which are combined in a final fusion stage. Each block is super- vised through a combination of supervised and gated self- supervision losses. To facilitate training and validation, we acquire a long-range synchronized gated stereo dataset for automotive scenarios. We find that the method achieves an improvement of more than 50 % MAE compared to the next best RGB stereo method, and 74 % MAE to existing monoc- ular gated methods for distances up to 160 m. Our code, models and datasets are available here1.	1. Introduction Long-range high-resolution depth estimation is critical for autonomous drones, robotics, and driver assistance sys- tems. Most existing fully autonomous vehicles strongly rely on scanning LiDAR for depth estimation [51, 52]. While these sensors are effective for obstacle avoidance the mea- surements are often not as semantically rich as RGB im- ages. LiDAR sensing also has to make trade-offs due to physical limitations, especially beyond 100 meters range, including range range versus eye-safety and spatial reso- lution. Although recent advances in LiDAR sensors such as, MEMS scanning [60] and photodiode technology [58] have drastically reduced the cost and led to a number of sen- sor designs with ≈100-200scanlines, these are still sig- nificantly lower resolutions than modern HDR megapixel camera sensors with a vertical resolution more than ≈5000 pixels. However, extracting depth from RGB images with monocular methods is challenging as existing estimation methods suffer from a fundamental scale ambiguity [16]. Stereo-based depth estimation methods resolve this issue but need to be well calibrated and often fail on texture-less 1https://light.princeton.edu/gatedstereo/regions and in low-light scenarios when no reliable features, and hence triangulation candidate, can be found. To overcome the limitations of existing scanning LiDAR and RGB stereo depth estimation methods, a body of work has explored gated imaging [2, 7–9, 22,27]. Gated im- agers integrate the transient response from flash-illuminated scenes in broad temporal bins, see Section 3for more de- tails. This imaging technique is robust to low-light, and adverse weather conditions [7] and the embedded time-of- flight information can be decoded as depth. Specifically, Gated2Depth [23] estimates depth from three gated slices and learns the prediction through a combination of simula- tion and LiDAR supervision. Building on these findings, re- cently, Walia et al. [59] proposed a self-supervised training approach predicting higher-quality depth maps. However, both methods have in common that they often fail in condi- tions where the signal-to-noise ratio is low, e.g., in the case of strong ambient light. We propose a depth estimation method from gated stereo observations that exploits both multi-view and time-of- flight cues to estimate high-resolution depth maps. We propose a depth reconstruction network that consists of a monocular depth network per gated camera and a stereo network that utilizes both active and passive slices from the gated stereo pair. The monocular network exploits depth- dependent gated intensity cues to estimate depth in monoc- ular and low-light regions while the stereo network relies on active stereo cues. Both network branches are fused in a learned fusion block. Using passive slices allows us to per- form robustly under bright daylight where active cues have a low signal-to-noise ratio due to ambient illumination. To train our network, we rely on supervised and self-supervised losses tailored to the stereo-gated setup, including ambient- aware and illuminator-aware consistency along with multi- camera consistency. To capture training data and assess the method, we built a custom prototype vehicle and captured a stereo-gated dataset under different lighting conditions and automotive driving scenarios in urban, suburban and high- way environments across 1000 km of driving. Specifically, we make the following contributions: • We propose a novel depth estimation approach us- ing gated stereo images that generates high-resolution This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13252 dense depth maps from multi-view and time-of-flight depth cues. • We introduce a depth estimation network with two different branches for depth estimation, a monocular branch and a stereo branch, that use active and passive measurement, and a semi-supervised training scheme to train the estimator. • We built a prototype vehicle to capture test and training data, allowing us to assess the method in long-range automotive scenes, where we reduce the MAE error by 50 % to the next best RGB stereo method and by 74 % on existing monocular gated methods for distances up to 160 m.
Wang_Improving_Robust_Generalization_by_Direct_PAC-Bayesian_Bound_Minimization_CVPR_2023	Abstract Recent research in robust optimization has shown an overﬁtting-like phenomenon in which models trained against adversarial attacks exhibit higher robustness on the training set compared to the test set. Although pre- vious work provided theoretical explanations for this phe- nomenon using a robust PAC-Bayesian bound over the ad- versarial test error, related algorithmic derivations are at best only loosely connected to this bound, which implies that there is still a gap between their empirical success and our understanding of adversarial robustness theory. To close this gap, in this paper we consider a different form of the robust PAC-Bayesian bound and directly min- imize it with respect to the model posterior. The derivation of the optimal solution connects PAC-Bayesian learning to the geometry of the robust loss surface through a Trace of Hessian (TrH) regularizer that measures the surface ﬂat- ness. In practice, we restrict the TrH regularizer to the top layer only, which results in an analytical solution to the bound whose computational cost does not depend on the network depth. Finally, we evaluate our TrH regulariza- tion approach over CIFAR-10/100 and ImageNet using Vi- sion Transformers (ViT) and compare against baseline ad- versarial robustness algorithms. Experimental results show that TrH regularization leads to improved ViT robustness that either matches or surpasses previous state-of-the-art approaches while at the same time requires less memory and computational cost.	1. Introduction Despite their success in a wide range of ﬁelds and tasks, deep learning models still remain susceptible to manipu- lating their outputs by even tiny perturbations to the input [6,7,10,19,30,32,43,47]. Several lines of work have fo- cused on developing robust training techniques against such *Work done in Google. †Corresponding authorLayer 1 Layer N-1Top Layer adversarialRobustness Loss !"cleanTrace of Hessian #$(∇!!"!")weights:(#weights:($+\|""\|##+\|"$\|##+Figure 1. We propose Trace of Hessian (TrH) regularization for training adversarially robust models. In addition to an ordinary robust loss (e.g., TRADES [ 54]), we regularize the TrH of the loss with respect to the weights of the top layer to encourage ﬂatness. The training objective is the result of direct PAC-Bayesian bound minimization in Theorem 3. adversarial attacks [ 8,20,31,32,36,41,46,48,54]. Im- portantly, Rice et al. [ 38] observe a robust overﬁtting phe- nomenon, referred to as the robust generalization gap , in which a robustly-trained classiﬁer shows much higher ac- curacy on adversarial examples from the training set, com- pared to lower accuracy on the test set. Indeed, several tech- nical approaches have been developed that could alleviate this overﬁtting phenomenon, including `2weight regular- ization, early stopping [ 38], label smoothing, data augmen- tation [ 51,53], using synthetic data [ 21] and etc. According to learning theory, the phenomenon of overﬁt- ting can be characterized by a PAC-Bayesian bound [ 4,9,18, 34,42] which upper-bounds the expected performance of a random classiﬁer over the underlying data distribution by its performance on a ﬁnite set of training points plus some ad- ditional terms. Although several prior works [ 21,24,26,49] have built upon insights from the PAC-Bayesian bound, none attempted to directly minimize the upper bound, likely due to the fact that the minimization of their forms of the PAC-Bayesian bound do not have an analytical solution. In this paper, we rely on a different form of the PAC- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16458 Bayesian bound [ 18], which can be readily optimized using a Gibbs distribution [ 16] with which we derive a second- order upper bound over the robust test loss. Interestingly, the resulting bound consists of a regularization term that in- volves Trace of Hessian (TrH) [ 12] of the network weights, a well-known measure of the loss-surface ﬂatness. For practical reasons, we limit TrH regularization to the top layer of the network only because computing a Hessian matrix and its trace for the entire network is too costly. We further derive the analytical expression of the top-layer TrH and show both theoretically and empirically that top-layer TrH regularization has a similar effect as regularizing the entire network. The resulting TrH regularization (illustrated in Figure 1) is less expensive and more memory efﬁcient compared to other competitive methods [ 21,24,26,49]. In summary, our contributions are as follows: (1) We provide a PAC-Bayesian upper-bound over the robust test loss and show how to directly minimize it (Theorem 3). To the best of our knowledge, this has not been done by prior work. Our bound includes a TrH term which encourages the model parameters to converge at a ﬂat area of the loss func- tion; (2) Taking efﬁciency into consideration, we restrict the TrH regularization to the top layer only (Algorithm 1) and show that it is an implicit but empirically effective reg- ularization on the TrH of each internal layer (Theorem 4 and Example 1); and (3) Finally, we conduct experiments with our new TrH regularization and compare the results to several baselines using Vision Transformers [ 14]. On CIFAR-10/100, our method consistently matches or beats the best baseline. On ImageNet, we report a signiﬁcant gain (+2.7%) in robust accuracy compared to the best baseline and establish a new state-of-the-art result of 48.9%.
Wang_Seeing_What_You_Said_Talking_Face_Generation_Guided_by_a_CVPR_2023	Abstract Talking face generation, also known as speech-to-lip generation, reconstructs facial motions concerning lips given coherent speech input. The previous studies revealed the importance of lip-speech synchronization and visual quality. Despite much progress, they hardly focus on the content of lip movements i.e., the visual intelligibility of the spoken words, which is an important aspect of generation quality. To address the problem, we propose using a lip- reading expert to improve the intelligibility of the gener- ated lip regions by penalizing the incorrect generation re- sults. Moreover, to compensate for data scarcity, we train the lip-reading expert in an audio-visual self-supervised manner. With a lip-reading expert, we propose a novel contrastive learning to enhance lip-speech synchronization, and a transformer to encode audio synchronically with video, while considering global temporal dependency of au- dio. For evaluation, we propose a new strategy with two different lip-reading experts to measure intelligibility of the generated videos. Rigorous experiments show that our pro- posal is superior to other State-of-the-art (SOTA) methods, such as Wav2Lip, in reading intelligibility i.e., over 38% Word Error Rate (WER) on LRS2 dataset and 27.8% ac- curacy on LRW dataset. We also achieve the SOTA perfor- mance in lip-speech synchronization and comparable per- formances in visual quality.	1. Introduction Talking Face Generation (TFG) aims at generating high- fidelity talking heads which are temporally synchronized with the input speech. It plays a significant role in many Human Robot Interaction (HRI) applications, such as film dubbing [1], video editing, face animation [2, 3], and com- munication with people who has hearing loss but master in lip-reading. Thanks to its various practical usage, TFG has ∗jiadong.wang@u.nus.edu †corresponding author (qianxy@ustb.edu.cn)also received an increasing attention in both industrial and research community over the past decades [4, 5]. In TFG, there are two major aspects of concerna: lip- speech synchronization and visual quality. While hu- mans are sensitive to subtle abnormalities in asynchronized speech and facial motions [6], the mechanism of speech production highly relies on lip movements [7]. As a result, the main challenge of TFG exists in temporal alignment between the input speech and synthesized video streams. One solution to this problem is to place an auxiliary em- bedding network at the end of the generator to analyze the audio-visual coherence. For example, [8] uses a pre-trained embedding network [9] as the lip-sync discriminator, while [10] investigates an asymmetric mutual information estima- tor. Another solution is to compute a sync loss between visual lip features of ground truth and generated video se- quences [11]. Other attempts use the encoder-decoder struc- ture to facilitate TFG by improving audio and visual repre- sentations, such as [12, 13], which disentangle the visual features to enhance audio representations into a shared la- tent space; or, [14], which disentangles audio factors, i.e., emotional and phonetic content, to remove sync-irrelevant features. Apart from lip-speech synchronization, blurry or unreal- istic visual quality also penalizes generation performance. To preserve defining facial features of target persons, skip connections [15] are applied [5]. Others [12, 16] employ Generative Adversarial Nets (GAN) to distinguish real and synthesized results by modelling the temporal dynamics of visual outputs. In this way, the resulting models can gen- erate more plausible talking heads that can be qualitatively measured by subjective evaluations. In addition, reading intelligibility should be indispens- able, but it has not been emphasized. Reading intelligibil- ity indicates how much text content can be interpreted from face videos by humans’ lip reading ability, which is espe- cially significant for hearing-impaired users. However, im- age quality and lip-speech synchronization do not explicitly reflect reading intelligibility. Specifically, a well-qualified This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14653 image may contain fine-grained lip-sync errors [8] while precise synchronization may convey incorrect text contents. According to the McGurk effect [17], when people listen and see an unpaired but synchronized sequence of speech and lip movements, they may recognize a phoneme from audio or video, or a fused artifact. In this paper, we propose a TalkLip net to synthesize talking faces by focusing on reading intelligibility. Specifi- cally, we employ a lip-reading expert which transcribes im- age sequences to text to penalize the generator for incor- rectly generated face images. We replace some images of a face sequence with the generated ones, and feed it to the lip-reading expert during training to supervise the face gen- erator. However, lip reading is hard even for humans. In [18], four people with equal gender distribution are invited to read lip movements. However, the average error rate is as high as 47%. Therefore, a reliable lip-reading model re- lies on a great amount of data. We employ A V-Hubert [19], a self-supervised method, which has yielded SOTA perfor- mance in lip-reading, speech recognition, and audio-visual speech recognition. The encoders of lip-reading and speech recognition systems are highly synchronized since they are supervised by the same pseudo label during pre-training. Leveraging the lip-reading expert from the A V-Hubert, we propose a new method to enhance lip-speech synchro- nization. Particularly, we conduct contrastive learning be- tween audio embeddings for face generation and visual con- text features from the lip-reading encoder. Besides, the A V-Hubert also provides a synchronized speech recognition system whose encoder considers long-term temporal depen- dency, we adopt this encoder to encode audio inputs, instead of encoders in [8, 11, 13] only rely on short-term temporal dependency (0.2s audio), or the single-modality (audio) pre- trained encoder in [20]. Our contributions are summarized as follows: • We tackle the reading intelligibility problem of speech- driven talking face generation by leveraging a lip- reading expert. • To enhance lip-speech synchronization, we propose a novel cross-modal contrastive learning strategy, as- sisted by a lip-reading expert. • We employ a transformer encoder trained synchron- ically with the lip-reading expert to consider global temporal dependency across the entire audio utterance. • We propose a new strategy to evaluate reading intelli- gibility for TFG and make the benchmark code pub- licly available . Code link: https://github.com/Sxjdwang/TalkLip• Extensive experiments demonstrate the feasibility of our proposal and its superiority over other prevail- ing methods in reading intelligibility (over 38% WER on LRS and 27.8% accuracy on LRW). Additionally, our approach performs comparably to or better than other SOTA methods in terms of visual quality and lip- speech synchronization.
Xiu_ECON_Explicit_Clothed_Humans_Optimized_via_Normal_Integration_CVPR_2023	Abstract The combination of deep learning, artist-curated scans, and Implicit Functions ( IF), is enabling the creation of de- tailed, clothed, 3D humans from images. However, existing methods are far from perfect. IF-based methods recover free-form geometry, but produce disembodied limbs or de- generate shapes for novel poses or clothes. To increase robustness for these cases, existing work uses an explicit parametric body model to constrain surface reconstruction, but this limits the recovery of free-form surfaces such as loose clothing that deviates from the body. What we want is a method that combines the best properties of implicit repre- sentation and explicit body regularization. To this end, we make two key observations: (1) current networks are better at inferring detailed 2D maps than full-3D surfaces, and (2) a parametric model can be seen as a “canvas” for stitch- ing together detailed surface patches. Based on these, our method, ECON , has three main steps: (1) It infers detailed 2D normal maps for the front and back side of a clothed per- son. (2) From these, it recovers 2.5D front and back surfaces, called d-BiNI , that are equally detailed, yet incomplete, and registers these w.r.t. each other with the help of a SMPL-Xbody mesh recovered from the image. (3) It “inpaints” the missing geometry between d-BiNI surfaces. If the face and hands are noisy, they can optionally be replaced with the ones of SMPL-X . As a result, ECON infers high-fidelity 3D humans even in loose clothes and challenging poses. This goes beyond previous methods, according to the quantitative evaluation on the CAPE andRenderpeople datasets. Per- ceptual studies also show that ECON ’s perceived realism is better by a large margin. Code and models are available for research purposes at econ.is.tue.mpg.de	1. Introduction Human avatars will be key for future games and movies, mixed-reality, tele-presence and the “metaverse”. To build re- alistic and personalized avatars at scale, we need to faithfully reconstruct detailed 3D humans from color photos taken in the wild. This is still an open problem, due to its challenges; people wear all kinds of different clothing and accessories, and they pose their bodies in many, often imaginative, ways. A good reconstruction method must accurately capture these, while also being robust to novel clothing and poses. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 512 Initial, promising, results have been made possible by using artist-curated scans as training data, and implicit func- tions ( IF) [56,59] as the 3D representation. Seminal work on PIFu(HD) [70, 71] uses “pixel-aligned” IFand reconstructs clothed 3D humans with unconstrained topology. However, these methods tend to overfit to the poses seen in the training data, and have no explicit knowledge about the human body’s structure. Consequently, they produce disembodied limbs or degenerate shapes for images with novel poses; see the 2nd row of Fig. 2. Follow-up work [26, 82, 96] accounts for such artifacts by regularizing the IFusing a shape prior provided by an explicit body model [52, 61], but regularization intro- duces a topological constraint, restricting generalization to novel clothing while attenuating shape details; see the 3rd and 4th rows of Fig. 2. In a nutshell, there are trade-offs between robustness, generalization and detail. What we want is the best of both worlds ; that is, the robustness of explicit anthropomorphic body models, and the flexibility of IFto capture arbitrary clothing topology. To that end, we make two key observations: (1) While inferring detailed 2D normal maps from color images is relatively easy [31, 71, 82], inferring 3D geometry with equally fine details is still challenging [9]. Thus, we exploit networks to infer detailed “geometry-aware” 2D maps that we then lift to 3D. (2) A body model can be seen as a low-frequency “canvas” that “guides” the stitching of detailed surface parts. With these in mind, we develop ECON , which stands for “Explicit Clothed humans Optimized via Normal inte- gration”. It takes, as input, an RGB image and a SMPL-X body inferred from the image. Then, it outputs a 3D human in free-form clothing with a level of detail and robustness that goes beyond the state of the art ( SOTA ); see the bottom of Fig. 2. Specifically, ECON has three steps . Step 1: Front & back normal reconstruction. We predict front- and back-side clothed-human normal maps from the input RGB image, conditioned on the body estimate, with a standard image-to-image translation network. Step 2: Front & back surface reconstruction. We take the previously predicted normal maps, and the correspond- ing depth maps that are rendered from the SMPL-X mesh, to produce detailed and coherent front-/back-side 3D sur- faces,{M F,MB}. To this end, we extend the recent BiNI method [7], and develop a novel optimization scheme that is aimed at satisfying three goals for the resulting surfaces: (1) their high-frequency components agree with clothed-human normals, (2) their low-frequency components and the dis- continuities agree with the SMPL-X ones, and (3) the depth values on their silhouettes are coherent with each other and consistent with the SMPL-X -based depth maps. The two out- put surfaces, {M F,MB}, are detailed yet incomplete, i.e., there is missing geometry in occluded and “profile” regions. Step 3: Full 3D shape completion. This module takes two inputs: (1) the SMPL-X mesh, and (2) the two d-BiNI Figure 2. Summary of SOTA .PIFuHD [71] recovers clothing details, but struggles with novel poses. ICON [82] and PaMIR [96] regularize shape to a body shape, but over-constrain the skirts, or over-smooth the wrinkles. ECON combines their best aspects. surfaces, {M F,MB}. The goal is to “inpaint” the missing geometry. Existing methods struggle with this problem. On one hand, Poisson reconstruction [38] produces “blobby” shapes and naively “infills” holes without exploiting a shape distribution prior. On the other hand, data-driven approaches, such as IF-Nets [10], struggle with missing parts caused by (self-)occlusions, and fail to keep the fine details present on two d-BiNI surfaces, producing degenerate geometries. We address above the limitations in two steps: (1) We ex- tend and re-train IF-Nets to be conditioned on the SMPL-X body, so that SMPL-X regularizes shape “infilling”. We dis- card the triangles that lie close to {M F,MB}, and keep the remaining ones as “infilling patches”. (2) We stitch together the front- and back-side surfaces and infilling patches via Poisson reconstruction; note that holes between these are small enough for a general purpose method. The result is a full 3D shape of a clothed human; see Fig. 2, bottom. We evaluate ECON both on established benchmarks (CAPE [55] and Renderpeople [66]) and in-the-wild images. Quantitative analysis reveals ECON ’s superiority. A percep- tual study echos this, showing that ECON is significantly preferred over competitors on challenging poses and loose clothing, and competitive with PIFuHD on fashion images. Qualitative results show that ECON generalizes better than theSOTA to a wide variety of poses and clothing, even with extreme looseness or complex topology; see Fig. 9. With both pose-robustness and topological flexibility, ECON recovers 3D clothed humans with a good level of detail and realistic pose. Code and models are available for research purposes at econ.is.tue.mpg.de 513
Wu_Cap4Video_What_Can_Auxiliary_Captions_Do_for_Text-Video_Retrieval_CVPR_2023	Abstract Most existing text-video retrieval methods focus on cross-modal matching between the visual content of videos and textual query sentences. However, in real-world sce- narios, online videos are often accompanied by relevant text information such as titles, tags, and even subtitles, which can be utilized to match textual queries. This insight has motivated us to propose a novel approach to text-video retrieval, where we directly generate associated captions from videos using zero-shot video captioning with knowl- edge from web-scale pre-trained models (e.g., CLIP and GPT-2). Given the generated captions, a natural ques- tion arises: what benefits do they bring to text-video re- trieval? To answer this, we introduce Cap4Video, a new framework that leverages captions in three ways: i) Input data: video-caption pairs can augment the training data. ii) Intermediate feature interaction: we perform cross-modal feature interaction between the video and caption to pro- duce enhanced video representations. iii) Output score: the Query-Caption matching branch can complement the original Query-Video matching branch for text-video re- trieval. We conduct comprehensive ablation studies to demonstrate the effectiveness of our approach. Without any post-processing, Cap4Video achieves state-of-the-art per- formance on four standard text-video retrieval benchmarks: MSR-VTT (51.4%), VATEX (66.6%), MSVD (51.8%), and DiDeMo (52.0%). The code is available at https:// github.com/whwu95/Cap4Video .	1. Introduction Text-video retrieval is a fundamental task in video- language learning. With the rapid advancements in image- language pre-training [15, 30, 46, 47], researchers have fo- cused on expanding pre-trained image-language models, es- pecially CLIP [30], to tackle the text-video retrieval task. The research path has evolved from the most direct global *Equal contribution. VideoCaptionsCLIPGPT-2 (a)Zero-Shot Caption Generation(b)CaptionsforVideoUnderstandingEnhancedRepresentationAugmentedTrainingPairsAuxiliaryMatchingScoresTextVideoVideoCaptions QueryCaption VideoQuery-VideoMatchingNewtrainingpairsQuery-CaptionMatchingCross-modalInteractionVideoQueryQuery-VideoMatchingVideoCaptionCLIPGPT-2Captioner(a)(b)(c) Figure 1. (a) An existing end-to-end learning paradigm for text- video retrieval. (b) Zero-shot video captioning achieved by guid- ing a large language model (LLM) such as GPT-2 [31] with CLIP [30]. (c) Our Cap4Video framework leverages the generated captions in three aspects: input data augmentation, intermediate feature interaction, and output score fusion. matching ( i.e., video-sentence alignment [11, 24]) to fine- grained matching ( e.g., frame-word alignment [36], video- word alignment [13], multi-hierarchical alignment [9, 28], etc.). These studies have demonstrated remarkable per- formance and significantly outperformed previous models. Two key factors contribute to this improvement. Firstly, CLIP offers powerful visual and textual representations that are pre-aligned in the semantic embedding space, thereby reducing the challenge of cross-modal learning in video- text matching. Secondly, these methods can fine-tune the pre-trained vision and text encoders using sparsely sampled frames in an end-to-end manner. All of these methods aim to learn cross-modal alignment between the visual represen- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10704 tation of videos and the textual representation of the corre- sponding query, as depicted in Figure 1(a). However, in real-life scenarios, online videos usually come with related content such as the video’s title or tag on the video website. In addition to the visual signal in the video, the associated textual information can also be used to some extent to describe the video content and match the query ( i.e., the common text-to-text retrieval). This raises a pertinent question: How can we generate associated text de- scriptions for videos? One possible solution is to crawl the video title from the video website. However, this method relies on annotations, and there is a risk that the video URL may have become invalid. Another automated solution is to generate captions using zero-shot video caption mod- els. Therefore, we turn our attention to knowledge-rich pre- trained models to handle such challenging
Wang_Practical_Network_Acceleration_With_Tiny_Sets_CVPR_2023	Abstract Due to data privacy issues, accelerating networks with tiny training sets has become a critical need in practice. Previous methods mainly adopt filter-level pruning to ac- celerate networks with scarce training samples. In this pa- per, we reveal that dropping blocks is a fundamentally su- perior approach in this scenario. It enjoys a higher ac- celeration ratio and results in a better latency-accuracy performance under the few-shot setting. To choose which blocks to drop, we propose a new concept namely recov- erability to measure the difficulty of recovering the com- pressed network. Our recoverability is efficient and effec- tive for choosing which blocks to drop. Finally, we propose an algorithm named PRACTISE to accelerate networks us- ing only tiny sets of training images. PRACTISE outper- forms previous methods by a significant margin. For 22% latency reduction, PRACTISE surpasses previous methods by on average 7% on ImageNet-1k. It also enjoys high generalization ability, working well under data-free or out- of-domain data settings, too. Our code is at https: //github.com/DoctorKey/Practise .	1. Introduction In recent years, convolutional neural networks (CNNs) have achieved remarkable success, but they suffer from high computational costs. To accelerate the networks, many net- work compression methods have been proposed, such as network pruning [11,18,20,22], network decoupling [6,15] and network quantization [2, 7]. However, most previous methods rely on the original training set (i.e., all the train- ing data) to recover the model’s accuracy. But, to preserve data privacy and/or to achieve fast deployment, only scarce training data may be available in many scenarios. For example, a customer often asks the algorithmic provider to accelerate their CNN models, but due to privacy *J. Wu is the corresponding author. This research was partly sup- ported by the National Natural Science Foundation of China under Grant 62276123 and Grant 61921006. /g1010/g1009/g1010/g1010/g1010/g1011/g1010/g1012/g1010/g1013/g1011/g1004/g1011/g1005/g1011/g1006/g1011/g1007/g1011/g1008 /g1006/g1012 /g1007/g1004 /g1007/g1006 /g1007/g1008 /g1007/g1010 /g1007/g1012 /g1008/g1004 /g1008/g1006/g100/g381/g393/g882/g1005/g3/g4/g272/g272/g856/g3/g894/g1081/g895 /g62/g258/g410/g286/g374/g272/g455/g3/g894/g373/g400/g895/g17/g367/g381/g272/g364 /g38/g349/g367/g410/g286/g396/g3/g894/g18/g24/g882/g400/g410/g455/g367/g286/g895 /g38/g349/g367/g410/g286/g396/g3/g894/g374/g381/g396/g373/g258/g367/g895 /g38/g349/g367/g410/g286/g396/g3/g894/g396/g286/g400/g349/g282/g437/g258/g367/g895 /g104/g374/g272/g381/g373/g393/g396/g286/g400/g400/g286/g282/g271/g286/g410/g410/g286/g396Figure 1. Comparison of different compression schemes with only 500 training images. We propose dropping blocks for few-shot network acceleration. Our method (‘Block’) outperforms previ- ous methods dominantly for the latency-accuracy tradeoff. The ResNet-34 model was compressed on ImageNet-1k and all laten- cies were tested on an NVIDIA TITAN Xp GPU. concerns, the whole training data cannot be available. Only the raw uncompressed model and a few training examples are presented to the algorithmic provider. In some extreme cases, not even a single data point is to be provided. The algorithmic engineers need to synthesize images or collect some out-of-domain training images by themselves. Hence, to learn or tune a deep learning model with only very few samples is emerging as a critical problem to be solved. In this few-shot compression scenario, most previous works [1,12,30] adopt filter-level pruning. However, it can- not achieve a high acceleration ratio on real-world com- puting devices (e.g., on GPUs). To make compressed mod- els indeed run faster than the uncompressed models, lots of FLOPs (number of floating point operations) are required to be reduced by filter-level pruning. And without the whole training dataset, it is difficult to recover the compressed model’s accuracy. Hence, previous few-shot compression This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20331 KD [10] FSKD [12] CD [1] MiR [30] BP (blocks) 44.5 45.3 56.2 64.1 66.5 Table 1. Top-1 validation accuracy (%) on ImageNet-1k for differ- ent compression schemes. ResNet-34 was accelerated by reducing 16% latency with 50 training images. Previous methods prune fil- ters with the ‘normal’ style. For the block-level pruning, we sim- ply remove the first kblocks and finetune the pruned network by back propagation, i.e., ‘BP (blocks)’ in this table. methods often exhibit a poor latency (wall-clock timing) vs. accuracy tradeoff. In this paper, we advocate that we need to focus on latency-accuracy rather than FLOPs-accuracy, and reveal that block-level pruning is fundamentally superior in the few-shot compression scenario. Compared to pruning fil- ters, dropping blocks enjoys a higher acceleration ratio. Therefore it can keep more capacity from the original model and its accuracy is easier to be recovered by a tiny train- ing set under the same latency when compared with filter pruning. Fig. 1 shows dropping blocks dominantly out- performs previous compression schemes for the latency- accuracy tradeoff. Table 1 further reports that an em- barrassingly simple dropping block baseline (i.e., finetune without any other processing) has already surpassed exist- ing methods which use complicated techniques . The base- line, ‘BP (blocks)’, simply removes the first few blocks and finetune the pruned network with the cross-entropy loss. To further improve block pruning, we study the strat- egy for choosing which blocks to drop, especially when only scarce training samples are available. Several cri- teria [21, 31, 34] have been proposed for pruning blocks on the whole dataset. However, some [31, 34] require a large amount of data for choosing, whereas others [21] only evaluate the output difference before/after block re- moval. In this paper, we notice that although dropping some blocks significantly changes the feature maps, they are easily recovered by end-to-end finetuning even with a tiny training set. So simply measuring the difference between pruned/original networks is not valid. To deal with these problems, a new concept namely recoverability is proposed in this paper for better indicating blocks to drop. And we propose a method to compute it efficiently, with only a few training images. At last, our recoverability is surprisingly consistent with the accuracy of the finetuned network. Finally, we propose P RACTISE , namely Pr actical net- work ac celeration with ti ny se ts of images, to effectively accelerate a network with scarce data. P RACTISE signifi- cantly outperforms previous few-shot pruning methods. For 22.1%latency reduction, P RACTISE surpasses the previous state-of-the-art (SOTA) method on average by 7.0% (per- centage points, not relative improvement) Top-1 accuracyon ImageNet-1k. It is also robust and enjoys high gener- alization ability which can be used on synthesized/out-of- domain images. Our contributions are: •We argue that the FLOPs-accuracy tradeoff is a mis- leading metric for few-shot compression, and advocate that the latency-accuracy tradeoff (which measures real runtime on devices) is more crucial in practice. For the first time, we find that in terms of latency vs. accuracy, block pruning is an embarrassingly simple but powerful method—dropping blocks with simple finetuning has already surpassed previ- ous methods (cf. Table 1). Note that although dropping blocks is previously known, we are the first to reveal its great potential in few-shot compression , which is both a sur- prising and an important finding. •To further boost the latency-accuracy performance of block pruning, we study the optimal strategy to drop blocks. A new concept recoverability is proposed to measure the difficulty of recovering each block, and in determining the priority to drop blocks. Then, we propose P RACTISE , an al- gorithm for accelerating networks with tiny sets of images. •Extensive experiments demonstrate the extraordinary performance of our P RACTISE . In both the few-shot and even the extreme data-free scenario, P RACTISE improves results by a significant margin. It is versatile and widely applicable for different network architectures, too.
Wang_Hard_Patches_Mining_for_Masked_Image_Modeling_CVPR_2023	Abstract Masked image modeling (MIM) has attracted much research attention due to its promising potential for learning scalable visual representations. In typical approaches, models usually focus on predicting specific contents of masked patches, and their performances are highly related to pre-defined mask strategies. Intuitively, this procedure can be considered as training a student (the model) on solving given problems (predict masked patches). However, we argue that the model should not only focus on solving given problems, but also stand in the shoes of a teacher to produce a more challenging problem by itself. To this end, we propose Hard Patches Mining (HPM), a brand-new framework for MIM pre-training. We observe that the reconstruction loss can naturally be the metric of the difficulty of the pre- training task. Therefore, we introduce an auxiliary loss predictor, predicting patch-wise losses first and deciding where to mask next. It adopts a relative relationship learning strategy to prevent overfitting to exact reconstruction loss values. Experiments under various settings demonstrate the effectiveness of HPM in constructing masked images. Furthermore, we empirically find that solely introducing the loss prediction objective leads to powerful representations, verifying the efficacy of the ability to be aware of where is hard to reconstruct.1	1. Introduction Self-supervised learning [6, 8, 9, 18, 20], with the goal of learning scalable feature representations from large- scale datasets without any annotations, has been a research hotspot in computer vision (CV). Inspired by masked 1Code: https://github.com/Haochen-Wang409/HPM Pre-definedmaskstrategy (a)ConventionalMIMpre-trainingparadigm. Whereishardtoreconstruct? imageinputmodelimage imageinputmodel(asastudent)image (b)OurproposedMIMpre-trainingparadigm.model(asateacher)Figure 1. Comparison between conventional MIM pre-training paradigm and our proposed HPM. (a)Conventional approaches can be interpreted as training a student , where the model is only equipped with the ability to solve a given problem under some pre-defined mask strategies. (b)Our proposed HPM pre-training paradigm makes the model to be both a teacher and a student , with the extra ability to produce a challenging pretext task . language modeling (MLM) [4,11,44,45] in natural language processing (NLP), where the model is urged to predict masked words within a sentence, masked image modeling (MIM), the counterpart in CV , has attracted numerous interests of researchers [3, 13, 19, 26, 42, 61, 66, 69]. Fig. 1a illustrates the paradigm of conventional ap- proaches for MIM pre-training [3, 19, 67]. In these typical solutions, models usually focus on predicting specific contents of masked patches. Intuitively, this procedure can be considered as training a student ( i.e., the model) on solving given problems ( i.e., predict masked patches). To alleviate the spatial redundancy in CV [19] and produce This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10375 Figure 2. Visual comparison between reconstruction loss anddiscriminativeness onImageNet validation set. We load the pre-trained ViT-B/16 [14] provided by MAE [19]. For each tuple, we show the (a)input image ,(b)patch-wise reconstruction loss averaged over 10 different masks, (c)predicted loss , and (d)masked images generated by the predicted loss ( i.e., patches with top 75% predicted loss are masked). Red means higher loss while blue indicates the opposite. Discriminative parts tend to be hard to reconstruct . a challenging pretext task, mask strategies become critical, which are usually generated under pre-defined manners, e.g., random masking [19], block-wise masking [3], and uniform masking [29]. However, we argue that a difficult pretext task is not all we need, and not only learning to solve the MIM problem is important, but also learning to produce challenging tasks is crucial. In other words, as shown in Fig. 1b, by learning to create challenging problems and solving them simultaneously , the model can stand in the shoes of both a student and a teacher , being forced to hold a more comprehensive understanding of the image contents, and thus leading itself by generating a more desirable task. To this end, we propose Hard Patches Mining (HPM) , a new training paradigm for MIM. Specifically, given an input image, instead of generating a binary mask under a manually- designed criterion, we first let the model be a teacher to produce a demanding mask, and then train the model to predict masked patches as a student just like conventional methods. Through this way, the model is urged to learn where it is worth being masked, and how to solve the problem at the same time. Then, the question becomes how to design the auxiliary task, to make the model aware of where the hard patches are. Intuitively, we observe that the reconstruction loss can be naturally a measure of the difficulty of the MIM task, which can be verified by the first two elements of each tuplein Fig. 2, where the backbone2pre-trained by MAE [19] with 1600 epochs is used for visualization. As expected, we find that those discriminative parts of an image ( e.g., object) are usually hard to reconstruct, resulting in larger losses. Therefore, by simply urging the model to predict reconstruction loss for each patch, and then masking those patches with higher predicted losses, we can obtain a more formidable MIM task. To achieve this, we introduce an auxiliary loss predictor, predicting patch-wise losses first and deciding where to mask next based on its outputs. To prevent it from being overwhelmed by the exact values of reconstruction losses and make it concentrate on the relative relationship among patches , we design a novel relative loss based on binary cross-entropy as the objective. We further evaluate the effectiveness of the loss predictor using a ViT-B under 200 epochs pre-training in Fig. 2. As the last two elements for each tuple in Fig. 2 suggest, patches with larger predicted losses tend to be discriminative, and thus masking these patches brings a challenging situation, where objects are almost masked. Meanwhile, considering the training evolution, we come up with an easy-to-hard mask generation strategy, providing some reasonable hints at the early stages. Empirically, we observe significant and consistent im- provements over the supervised baseline and vanilla MIM 2https://dl.fbaipublicfiles.com/mae/visualize/ mae_visualize_vit_base.pth 10376 pre-training under various settings. Concretely, with only 800 epochs pre-training, HPM achieves 84.2% and 85.8% Top-1 accuracy on ImageNet-1K [49] using ViT-B and ViT- L, outperforming MAE [19] pre-trained with 1600 epochs by +0.6% and +0.7%, respectively.
Wang_FeatureBooster_Boosting_Feature_Descriptors_With_a_Lightweight_Neural_Network_CVPR_2023	Abstract We introduce a lightweight network to improve descrip- tors of keypoints within the same image. The network takes the original descriptors and the geometric properties of key-points as the input, and uses an MLP-based self-boosting stage and a Transformer-based cross-boosting stage to en-hance the descriptors. The boosted descriptors can be ei- ther real-valued or binary ones. We use the proposed net- work to boost both hand-crafted (ORB [ 34], SIFT [ 24]) and the state-of-the-art learning-based descriptors (SuperPoint [10], ALIKE [ 53]) and evaluate them on image matching, visual localization, and structure-from-motion tasks. Theresults show that our method signiﬁcantly improves the per-formance of each task, particularly in challenging cases such as large illumination changes or repetitive patterns. Our method requires only 3.2ms on desktop GPU and 27mson embedded GPU to process 2000 features, which is fast enough to be applied to a practical system. The code and trained weights are publicly available at github.com/SJTU- ViSYS/FeatureBooster .	1. Introduction Extracting sparse keypoints or local features from an im- age is a fundamental building block in various computer vi- sion tasks, such as structure from motion (SfM), simultane- ous localization and mapping (SLAM), and visual localiza-tion. The feature descriptor, represented by a real-valued or binary descriptor, plays a key role in matching those key-points across different images. The descriptors are commonly hand-crafted in the early days. Recently, learning-based descriptors [ 10,53]h a v e *Corresponding Author: Danping Zou ( dpzou@sjtu.edu.cn ). This works was supported by National Key R&D Program(2022YFB3903802) and National of Science Foundation of China(62073214) Figure 1. ORB descriptors perform remarkably better in challeng- ing cases after being boosted by the proposed lightweight network. Left column : Matching results of using raw ORB descriptors. Right column : Results of using boosted ORB descriptors. Near- est neighbor search and RANSAC [ 14] were used for matching. shown to be more powerful than hand-crafted ones, espe- cially in challenging cases such as signiﬁcant viewpoint and illumination changes. Both hand-crafted and learning- based descriptors have shown to work well in practice.Some of them have become default descriptors for some ap-plications. For example, the simple binary descriptor ORB [34] is widely used for SLAM systems [ 20,29]. SIFT [ 24] is typically used in structure-from-motion systems. Considering that the descriptors have already been inte- grated into practical systems, replacing them with totallynew ones can be problematic, as it may require more com-puting power that may not be supported by the existinghardware, or sometimes require extensive modiﬁcations to the software because of changed descriptor type ( e.g. from binary to real). In this work, we attempt to reuse existing descriptors and enhance their discrimination ability with as little com- putational overhead as possible. To this end, we propose a lightweight network to improve the original descriptors. The input of this network is the descriptors and the geomet- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7630 ric properties such as the 2D locations of all the keypoints within the entire image. Each descriptor is ﬁrstly processedby an MLP (Multi-layer perceptron) and summed with ge-ometric properties encoded by another MLP . The new ge-ometrically encoded descriptors are then aggregated by an efﬁcient Transformer to produce powerful descriptors that are aware of the high-level visual context and spatial lay-out of those keypoints. The enhanced descriptors can beeither real-valued or binary ones and matched by using Eu-clidean/Hamming distance respectively. The core idea of our approach, motivated by recent work [25,36,41], is integrating the visual and geometric infor- mation of all the keypoints into individual descriptors by a Transformer. This can be better understood intuitively by considering when people are asked to ﬁnd correspondences between images, they would check all the keypoints and the spatial layout of those keypoints in each image. With the help of the global receptive ﬁeld in Transformer, the boosted descriptors contain global contextual informationthat makes them more robust and discriminative as shown in Fig. 1. We apply our FeatureBooster to both hand-crafted de- scriptors (SIFT [ 24], ORB [ 34]) and the state-of-the-art learning-based descriptors (SuperPoint [ 10], ALIKE [ 53]). We evaluated the boosted descriptors on tasks includingimage matching, visual localization, and structure-from-motion. The results show that our method can signiﬁcantlyimprove the performance of each task by using our boosted descriptors. Because FeatureBooster does not need to process the im- age and adopts a lightweight Transformer, it is highly efﬁ- cient. It takes only 3.2ms on NVIDIA RTX 3090 and 27mson NVIDIA Jetson Xavier NX (for embedded devices) toboost 2000 features, which makes our method applicable topractical systems.
Wang_Image_Cropping_With_Spatial-Aware_Feature_and_Rank_Consistency_CVPR_2023	Abstract Image cropping aims to find visually appealing crops in an image. Despite the great progress made by previous methods, they are weak in capturing the spatial relationship between crops and aesthetic elements ( e.g., salient objects, semantic edges). Besides, due to the high annotation cost of labeled data, the potential of unlabeled data awaits to be excavated. To address the first issue, we propose spatial- aware feature to encode the spatial relationship between candidate crops and aesthetic elements, by feeding the con- catenation of crop mask and selectively aggregated feature maps to a light-weighted encoder. To address the second issue, we train a pair-wise ranking classifier on labeled images and transfer such knowledge to unlabeled images to enforce rank consistency. Experimental results on the benchmark datasets show that our proposed method per- forms favorably against state-of-the-art methods.	1. Introduction The task of image cropping aims to find good crops in an image that can improve the image quality and meet aes- thetic requirement. Image cropping is a prevalent and criti- cal operation in numerous photography-related applications like image thumbnailing, view recommendation, and cam- era view adjustment suggestion. Many Researchers [2, 4–7, 12, 21, 23, 36, 43, 46, 52, 54, 60, 62, 63] have studied automatic image cropping in the past decades with the goal to reduce the workload of man- ual cropping. Earlier works [2, 3, 12, 31, 43, 44] mainly used saliency detection [49, 59] to detect salient objects and crop around salient objects. Another group of meth- ods [6, 12, 26, 33, 54, 62] designed hand-crafted features to represent specific composition rules in photography. With the construction of moderate-sized image cropping datasets [4, 52, 54, 56], recently proposed image cropping methods [4, 5, 7, 21, 23, 36, 52, 56, 57, 63] are usually data- driven manner and directly learn how to crop visually ap- *Corresponding author Source Image Low-level High-level Figure 1. Two examples of the spatial relationship between crops (yellow bounding box) and aesthetic elements ( e.g., semantic edges and salient objects). The first column shows the source im- ages, and the second ( resp. , third) column shows their low-level (resp. , high-level) feature maps extracted by a pre-trained Mo- bileNetv2 [39] network with channel-wise max pooling. It can be seen that low-level feature maps emphasize semantic edges and high-level feature maps highlight salient objects. pealing views from the labeled data. Although these ap- proaches have achieved impressive improvement on image cropping task, there still exist some drawbacks which will be discussed below. One problem is that when considering the spatial rela- tionship between crops and aesthetic elements ( e.g., salient objects, semantic edges), which is very critical for image cropping, previous methods usually designed some intu- itive rules. For example, the crop should enclose the salient object [2, 43, 44], or should not cut through the semantic edges [2, 54]. However, these hand-crafted rules did not consider the spatial layout of all aesthetic elements as a whole, and may not generalize well to various scenes be- cause the rules designed for specific subjects can not cover complex image cropping principles [10]. In this work, we explore learnable spatial-aware fea- tures, which encode the spatial relationship between crops and aesthetic elements. We observe that the feature map obtained using channel-wise max pooling can emphasize some aesthetic elements. In Figure 1, we show several pooled feature maps from MobileNetv2 [39], from which This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10052 it can be seen that the low-level feature maps emphasize semantic edges ( e.g., the outlines of semantic objects and regions) and the high-level feature maps emphasize salient objects ( e.g., bird, balloon). With concatenated feature maps from different layers, we learn channel attention [16] to select important layers. The weighted feature maps are concatenated with candidate crop masks and sent to a light- weighted encoder to produce spatial-aware features. The extracted spatial-aware features encode the spatial relation- ship between candidate crops and aesthetic elements with- out being limited by any hand-crafted rules. Another problem is that the cost of crop annotation is very high and the performance is limited by the scale of the annotated training set. Therefore, some previous works explored how to utilize unlabeled data to improve the crop- ping performance. For example, VFN [5] collects unlabeled professional photographs from public websites and perform pairwise ranking based on the assumption that the entire image has higher aesthetic quality than any of its crops. However, such assumption does not always hold obviously. VPN [52] used a pre-trained network VEN [52] to predict aesthetic scores for the crops from unlabeled images, which function as pseudo labels to supervise training a new net- work. However, the predicted pseudo labels may be very noisy and provide misleading guidance. In this work, we explore transferring ranking knowl- edge from labeled images to unlabeled images. Specifically, given two annotated crops from a labeled image, we learn a binary pairwise ranking classifier to judge which crop has higher aesthetic quality, by sending the concatenation of two crop features to a fully connected layer. We expect that the knowledge of comparing the aesthetic quality of two crops with similar content could be transferred to unlabeled data. Given two unannotated crops from an unlabeled im- age, we can obtain two types of ranks. On the one hand, we can rank them according to the predicted crop-level scores. On the other hand, we can employ the pairwise ranking clas- sifier to get the rank. Then, we enforce two types of ranks to be consistent. We conduct experiments on GAICD [57] and FCDB [4] dataset. For unlabeled images, we use unlabeled test im- ages, which falls into the scope of transductive learning. Our major contributions can be summarized as: • We design a novel spatial-aware feature to model the spatial relationship between candidate crops and aes- thetic elements. • We propose to transfer ranking knowledge from la- beled images to unlabeled images, and enforce ranking consistency on unlabeled images. • Our proposed method obtains the state-of-the-art per- formance on benchmark datasets.
Villa_PIVOT_Prompting_for_Video_Continual_Learning_CVPR_2023	Abstract Modern machine learning pipelines are limited due to data availability, storage quotas, privacy regulations, and expensive annotation processes. These constraints make it difficult or impossible to train and update large-scale mod- els on such dynamic annotated sets. Continual learning di- rectly approaches this problem, with the ultimate goal of devising methods where a deep neural network effectively learns relevant patterns for new (unseen) classes, without significantly altering its performance on previously learned ones. In this paper, we address the problem of contin- ual learning for video data. We introduce PIVOT, a novel method that leverages extensive knowledge in pre-trained models from the image domain, thereby reducing the num- ber of trainable parameters and the associated forgetting. Unlike previous methods, ours is the first approach that ef- fectively uses prompting mechanisms for continual learning without any in-domain pre-training. Our experiments show that PIVOT improves state-of-the-art methods by a signifi- cant 27% on the 20-task ActivityNet setup.	1. Introduction Modern Deep Neural Networks (DNNs) are at the core of many state-of-the-art methods in machine vision [10, 20, 21, 38, 44, 46] tasks. To achieve their remarkable performance, most DNNs rely on large-scale pre-training [11, 26, 36, 46], thereby enabling feature reuse in related downstream tasks [4]. However, adapting and fine-tuning a pre-trained DNN on a novel dataset commonly leads to catastrophic forgetting [16]. This phenomenon explains how the effectiveness of the fine-tuned DNNs drastically reduces in the original training distribution, in favor of in- creased performance on the downstream task. This undesirable property has driven multiple research efforts in the area of Continual Learning (CL) [7, 8, 18, 25, Figure 1. Performance improvement by each PIVOT com- ponent. We report the average accuracy on all tasks under the 10-task CIL on UCF101 and ActivityNet. We report the per- formance of basic CLIP, and then gradually equip it with other components: Spatial Prompting, Memory Buffer, Multi-modal Contrastive Learning, Temporal Encoder, to finally reach our pro- posed PIVOT method. The addition of each proposed component generally boosts the performance on both datasets. Stars represent the upper bound performance on each benchmark. 29], yielding techniques that enable fine-tuning a DNN on a sequence of tasks while mitigating the performance drop along the intermediate steps. One of the most challenging scenarios for the study of CL is Class Incremental Learning (CIL), where the labels and data are mutually exclusive be- tween tasks, training data is available only for the current task, and there are no task identifiers on the validation step (i.etask boundaries are not available at test time). Such a setup requires learning one model that, despite the continu- ous adaptation to novel tasks, performs well on all the seen classes. Recent advances in mitigating catastrophic forget- ting rely on deploying episodic memory or regularization techniques [2,8,25,37]. Nevertheless, most of this progress has been directed toward analyzing the catastrophic forget- ting of DNNs in the image domain. The works of [34,35,47] introduced the CIL setup to the video domain, in particular the action recognition task. Unlike its image counterpart, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 24214 video CIL requires careful modeling of the temporal infor- mation, making it an even more challenging setup. Despite the success in small-scale datasets (like UCF101), state-of- the-art methods have shown limited effectiveness on more challenging video test-beds built upon larger action tax- onomies, such as Kinetics and ActivityNet [47]. Currently, state-of-the-art methods for video CIL rely on temporal masking and feature distillation to mitigate catas- trophic forgetting [34,47]. In this paper, we take inspiration from recent advances in large-scale DNNs for zero-shot im- age classification [36] and learnable prompts for continual learning [50], and we propose a novel strategy for CIL in the video domain. We show that a zero-shot baseline pre- trained in the image domain already outperforms the best CIL methods in the action recognition task *. Moreover, we show that this baseline can be significantly improved by enabling temporal reasoning and augmenting the modal- ity encoders with a novel prompting strategy. As a con- sequence, our proposed method, PIVOT, outperforms ev- ery other baseline, setting a new state-of-the-art in all the 3 datasets included in the challenging vCLIMB benchmark for video CIL [47]. Figure 1 summarizes the performance improvements of our approach. Notably and following the core ideas of the vCLIMB [47] benchmark, PIVOT does not rely on any in-distribution pre-training (a common feature of prompting methods for CL [45,50]). Rather, it leverages the vast and general visual knowledge contained in the CLIP visual encoder (trained on massive amounts of paired static images and text) and maps that knowledge into a feature space suitable for video understanding in a continual learning setup. Contributions. This paper proposes PIVOT (Prompt- Ing for Video cOnTinual learning), a novel strategy for con- tinual learning in the video domains that leverages large- scale pre-trained networks in the image domain. Our work brings the following contributions: (i)We show that a multi- modal classifier (Video-Text) mitigates catastrophic forget- ting while greatly increasing the final average CIL accuracy. (ii)We design the first prompt-based strategy for video CIL. Our approach leverages image pre-training to significantly mitigate forgetting when learning a sequence of video ac- tion recognition tasks. (iii)We conduct extensive experi- mental analysis to demonstrate the effectiveness of PIVOT. Our results show that PIVOT outperforms state-of-the-art methods in the challenging vCLIMB benchmark by 31%, 27%, and 17.2% in the 20-task setups of Kinetics, Activi- tyNet, and UCF101, respectively.
Wang_Binary_Latent_Diffusion_CVPR_2023	Abstract In this paper, we show that a binary latent space can be explored for compact yet expressive image representa- tions. We model the bi-directional mappings between an image and the corresponding latent binary representation by training an auto-encoder with a Bernoulli encoding dis- tribution. On the one hand, the binary latent space provides a compact discrete image representation of which the distri- bution can be modeled more efﬁciently than pixels or con- tinuous latent representations. On the other hand, we now represent each image patch as a binary vector instead of an index of a learned cookbook as in discrete image repre- sentations with vector quantization. In this way, we obtain binary latent representations that allow for better image quality and high-resolution image representations without any multi-stage hierarchy in the latent space. In this binary latent space, images can now be generated effectively us- ing a binary latent diffusion model tailored speciﬁcally for modeling the prior over the binary image representations. We present both conditional and unconditional image gen- eration experiments with multiple datasets, and show that the proposed method performs comparably to state-of-the- art methods while dramatically improving the sampling ef- ﬁciency to as few as 16 steps without using any test-time acceleration. The proposed framework can also be seam- lessly scaled to 1024 ⇥1024 high-resolution image gener- ation without resorting to latent hierarchy or multi-stage reﬁnements.	1. Introduction The goal of modeling the image distribution that allows the efﬁcient generation of high-quality novel samples drives the research of representation learning and generative mod- els. Directly representing and generating images in the pixel space stimulates various research such as generative adver- sarial networks [ 2,7,18,29], ﬂow models [ 11,34,43,47], energy-based models [ 12,13,66,71], and diffusion mod- els [24,41,54,55]. As the resolution grows, it becomes increasingly difﬁcult to accurately regress the pixel values. And this challenge usually has to be addressed through hi- FFHQCelebA-HQ LSUNChurches 1024x1024256x256LSUNBedroomsImageNet FFHQCelebA Figure 1. Examples of generated images with different resolutions using the proposed method. erarchical model architectures [ 29,72] or at a notably high cost [ 24]. Moreover, while demonstrating outstanding gen- erated image quality, GAN models suffer from issues in- cluding insufﬁcient mode coverage [ 38] and training insta- bility [ 21]. Representing and generating images in a learned latent space [ 33,42,49] provides a promising alternative. Latent diffusion [ 49] performs denoising in latent feature space with a lower dimension than the pixel space, therefore re- ducing the cost of each denoising step. However, regress- ing the real-value latent representations remains complex and demands hundreds of diffusion steps. Variational auto- encoders (V AEs) [ 23,33,48] generate images without any it- erative steps. However, the static prior of the latent space re- stricts the expressiveness, and can lead to posterior collapse. To achieve higher ﬂexibility of the latent distribution with- out signiﬁcantly increasing the modeling complexity, VQ- V AE [ 61] introduces a vector-quantized latent space, where This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22576 each image is represented as a sequence of indexes, each of which points to a vector in a learned codebook. The prior over the vector-quantized representations is then modeled by a trained sampler, which is usually parametrized as an autoregressive model. The success of VQ-V AE stimulates a series of works that model the discrete latent space of code- book indexes with different models such as accelerated par- allel autoregressive models [ 8] and multinomial diffusion models [ 6,20]. VQ-based generative models demonstrate surprising image synthesis performance and model cover- age that is better than the more sophisticated methods like GANs without suffering from issues like training instability. However, the hard restriction of using one codebook index to represent each image patch introduces a trade-off on the codebook size, as a large enough codebook to cover more image patterns will introduce an over-complex multinomial latent distribution for the sampler to model. In this research, we explore a compact yet expressive representation of images in a binary latent space, where each image patch is now represented as a binary vector, and the prior over the discrete binary latent codes is effectively modeled by our improved binary diffusion model tailored for Bernoulli distribution. Speciﬁcally, the bi-directional mappings between images and the binary representations are modeled by a feed-forward autoencoder with a binary latent space. Given an image, the encoder now outputs the normalized parameters of a sequence of independently dis- tributed Bernoulli variables, from which a binary represen- tation of this image is sampled, and fed into the decoder to reconstruct the image. The discrete sampling in Bernoulli distribution does not naturally permit gradient propagation. We ﬁnd that a simple straight-through gradient copy [ 4,17] is sufﬁcient for high-quality image reconstruction while maintaining high training efﬁciency. With images compactly represented in the binary latent space, we then introduce how to generate novel samples by modeling the prior over binary latent codes of images. To overcome the shortcomings of many existing generative models such as being uni-directional [ 46,61] and the non- regrettable greedy sampling [ 6,8], we introduce binary la- tent diffusion that generates the binary representations of novel samples by a sequence of denoising starting from a random Bernoulli distribution. Performing diffusion in a binary latent space, modeled as Bernoulli distribution, re- duces the need for precisely regressing the target values as in Gaussian-based diffusion processes [ 24,49,55], and per- mits sampling at a higher efﬁciency. We then introduce how to progressively reparametrize the prediction targets at each denoising step as the residual between the inputs and the de- sired samples, and train the proposed binary latent diffusion models to predict such ‘ﬂipping probability’ for improved training and sampling stability. We support our ﬁndings with both conditional and unconditional image generation experiments on multiple datasets. We show that our method can deliver remarkable image generation quality and diversity with more compactlatent codes, larger image-to-latent resolution ratios, as well as fewer sampling steps, and faster sampling speed. We present some examples with different resolutions generated by the proposed method in Figure 1. We organize this paper as follows: Related works are discussed in Section 2. In Section 3, we introduce binary image representations by training an auto-encoder with a bi- nary latent space. We then introduce in Section 4binary la- tent diffusion, a diffusion model for multi-variate Bernoulli distribution, and techniques tailored speciﬁcally for improv- ing the training and sampling of binary latent diffusion. We present both quantitative and qualitative experimental re- sults in Section 5and conclude the paper in Section 6.
Wei_Physically_Adversarial_Infrared_Patches_With_Learnable_Shapes_and_Locations_CVPR_2023	Abstract Owing to the extensive application of infrared object de- tectors in the safety-critical tasks, it is necessary to evaluate their robustness against adversarial examples in the real world. However, current few physical infrared attacks are complicated to implement in practical application because of their complex transformation from digital world to physi- cal world. To address this issue, in this paper, we propose a physically feasible infrared attack method called ”adversar- ial infrared patches”. Considering the imaging mechanism of infrared cameras by capturing objects’ thermal radiation, adversarial infrared patches conduct attacks by attaching a patch of thermal insulation materials on the target object to manipulate its thermal distribution. To enhance adversarial attacks, we present a novel aggregation regularization to guide the simultaneous learning for the patch’ shape and location on the target object. Thus, a simple gradient-based optimization can be adapted to solve for them. We verify adversarial infrared patches in different object detection tasks with various object detectors. Experimental results show that our method achieves more than 90% Attack Suc- cess Rate (ASR) versus the pedestrian detector and vehicle detector in the physical environment, where the objects are captured in different angles, distances, postures, and scenes. More importantly, adversarial infrared patch is easy to im- plement, and it only needs 0.5 hours to be constructed in the physical world, which verifies its effectiveness and efficiency.	1. Introduction Deep Neural Networks (DNNs) have shown promising performance in various vision tasks, including object de- tection [14], classification [8], face recognition [16], and autonomous driving [15]. However, it is typically known that DNNs are vulnerable to adversarial examples [2, 5, 26], i.e., the human-imperceptible perturbed inputs can fool the DNNs-based system to give wrong predictions. Moreover, *Corresponding author Original Object Perturbed ObjectPatch in physical worldattackPatch in digital world stickFigure 1. The generation process of adversarial infrared patches. We see the pedestrian cannot be detected after the infrared patches are pasted on the pedestrian in the physical world. these adversarial examples can be exploited in the physical world. In such cases, a widely used technique is called ad- versarial patches [1, 4, 24, 28], which have been successfully applied to traffic sign detection by generating a carefully designed sticker [4, 23], or face recognition by adding spe- cific textures on eyeglass frames [17, 22]. The success of adversarial patches raises the concerns because of their great threat to the deployed DNN-based systems in the real world. Nowadays, famous for its strong anti-interference ability in the severe environment, object detection in the thermal infrared images has been widely used in many safety-critical tasks such as security surveillance [19], remote sensing [25], etc. Consequently, it is necessary to evaluate the physical adversarial robustness of infrared object detectors. However, the aforementioned adversarial patches cannot work well in the infrared images because they depend on the adversarial perturbations generated from the view of RGB appearance. These perturbations cannot be captured by infrared cameras, which perform the imaging by encoding the objects’ ther- mal radiation [20]. Although few recent works have been proposed to address this issue, they have their own limita- tions. For example, adversarial bulbs based on a cardboard of alight small bulbs [30] are complicated to implement in the real world, and are also not stealthy because they produce heat source. Adversarial clothing [29] based on a large-scale QR code improves the stealthiness by utilizing the thermal insulation material to cover the object’s thermal radiation, but still has complex transformations from digital to physical world, making it not easy to implement in the real world. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12334 Figure 2. The comparison between different infrared attacks. In this paper, we propose a physically stealthy and easy- to-implement infrared attack method called “adversarial in- frared patches”. Considering the imaging mechanism of infrared cameras by capturing objects’ thermal radiation, we also attach the thermal insulation materials on the target object to manipulate its thermal distribution to ensure the stealthiness. But different from adversarial clothing [29] via the complex QR code pattern, we utilize a simple patch to crop the thermal insulation material, and then adjust the patch’s shape and location on the target object to conduct attacks. Compared with adversarial RGB perturbations, the changes of shapes and locations of the thermal patch can be accurately captured by infrared cameras, which helps perform an effective attack. However, the shape and location are two kinds of different variables, it is challenging to directly optimize them via unified gradients. For that, we present a novel aggregation regularization to guide the simultaneous learning for the patch’ shape and location on the target object. Specifically, an aggregation degree is proposed to quantify how close one pixel’s neighbours are to being a clique. By combining this metric with the attack goal, the object’s pixels needing to be covered by the thermal insulation material will automatically be gathered to form a valid shape on the available location of the object. In this way, we can adapt a simple gradient-based optimization to solve for the optimal shape and location of the patch. An example of our adversarial infrared patch against the pedestrian detector is shown in Figure 1, and a comparison with the existing physical infrared attacks is given in Figure 2. We see that adversarial infrared patch is simpler than other methods in the digital world, and we just need to crop the thermal insulation materials according to the learned shape, and then paste the patch on the learned location of the pedestrian, which is more easy-to-implement than other methods in the real world. Our contributions can be summarized as follows: •We propose the novel “adversarial infrared patches”, a physically stealthy and easy-to-implement attack method for infrared object detection. Instead of gen- erating adversarial perturbations, we perform attacks by learning the available shape and location on the tar-get object. Owing to this careful design, adversarial infrared patches are easier to implement in the physical world than the existing methods. •We design a novel aggregation regularization to guide the simultaneous learning for the patch’ shape and loca- tion on the target object. Thus, a simple gradient-based optimization can be adapted to solve for the optimal shape and location of the patch. •We verify the adversarial infrared patches in the pedes- trian detection task from both the digital and physi- cal world. Experiments show that adversarial infrared patches can work well in various angles, distances, pos- tures, and scenes, and achieve competitive attacking performance with the SOTA infrared attack while only costing five percent of their time to construct physical adversarial examples. We also extend our method to the vehicle detection task to verify its generalization.
Wang_Raw_Image_Reconstruction_With_Learned_Compact_Metadata_CVPR_2023	Abstract While raw images exhibit advantages over sRGB images (e.g., linearity and fine-grained quantization level), they are not widely used by common users due to the large stor- age requirements. Very recent works propose to compress raw images by designing the sampling masks in the raw image pixel space, leading to suboptimal image represen- tations and redundant metadata. In this paper, we propose a novel framework to learn a compact representation in the latent space serving as the metadata in an end-to-end man- ner. Furthermore, we propose a novel sRGB-guided con- text model with the improved entropy estimation strategies, which leads to better reconstruction quality, smaller size of metadata, and faster speed. We illustrate how the pro- posed raw image compression scheme can adaptively al- locate more bits to image regions that are important from a global perspective. The experimental results show that the proposed method can achieve superior raw image re- construction results using a smaller size of the metadata on both uncompressed sRGB images and JPEG images. The code will be released at https://github.com/ wyf0912/R2LCM .	1. Introduction As an unprocessed and uncompressed data format di- rectly obtained from camera sensors, raw images has unique advantages for computer vision tasks in practice. For exam- ple, it is easier to model the distribution of real image noise in raw space, which enables generalized deep real denoising networks [1, 40]; As pixel values in raw images have a lin- ear relationship with scene radiance, they own benefits to re- cover shadows and highlights without bringing in the grainy noise usually associated with high ISO [12, 16, 33, 35], which greatly contributes to the low-light image enhance- ment. Besides, with richer colors, raw images offer more room for correction and artistic manipulation. *Corresponding author. UNetISPRawsRGBSampledRawPixelssRGBRawDecoderReconstructedRawSamplingmask(a) Previous SOTA methods sample in the raw space [26, 30]. Learned CompactFeatureEncoderDecodersRGBReconstructedRawRaw bitstream01···10Learnedentropycoding (b) The proposed method samples in the latent space. Figure 1. The comparison between the previous SOTA methods (in blue) and our proposed method (in green). Different from the pre- vious work where the sampling strategy is hand-crafted or learned by a pre-defined sampling loss, we learn the sampling and recon- struction process in a unified end-to-end manner. In addition, the sampling of previous works is in the raw pixel space, which in fact still includes a large amount of spatial redundancy and precision redundancy. Instead, we conduct sampling in the feature space, and more compact metadata is obtained for pixels in the feature space via the adaptive allocation. The saved metadata is annotated in dashed box. Despite of these merits, raw images are not widely adopted by common users due to large file sizes. In addi- tion, since raw images are unprocessed, additional post pro- cessing steps, e.g., demosaicing and denoising, are always needed before displaying them. For fast image rendering in practice, a copy of JPEG image is usually saved along with its raw data [2]. To improve the storage efficiency, raw- image reconstruction problem attracts more and more atten- tion, i.e., how to minimize the amount of metadata required for de-rendering sRGB images back to raw space. Classic metadata-based raw image reconstruction methods model the workflow of image signal processing (ISP) pipeline and save the required parameters in ISP as metadata [27]. To further reduce the storage and computational complexity to- wards a lightweight and flexible reverse ISP reconstruction, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18206 DynamicrangeclippingQuantizationNon-linearmapping(Bothgloballyandlocally)(a) Simplified ISP adopted from [21] (b) Processed Raw (c) Quantized sRGB (d) Quantization error map Figure 2. An illustration of the information loss caused by the ISP. (a) A simplified ISP suffers from the information loss caused by nonlinear transformations. (b) Raw image after process to better display the details. (c) Quantized sRGB image after ISP which suffers information loss, e.g., the red bounding box area. (d) The quantization error map. As we can see from the above figures, the information loss caused by the quantization is non-uniformly dis- tributed in both over-exposed areas and normally-exposed areas. very recent methods focus on sparse sampling of raw image pixels [26, 30]. Specifically, in [30], a uniform sampling strategy is proposed to combine with an interpolation al- gorithm that solves systems of linear equations. The work in [26] proposes a sampling network and approximates the reconstruction process by deep learning to further improve the sampling strategy. Though lots of progress has been made, existing sparse sampling based raw image reconstruction methods still face limitations in terms of coding efficiency and image recon- struction quality. Specifically, the bit allocation should be adaptive and globally optimized for the image contents, given the non-linear transformation and quantization steps in ISP as shown in Fig. 2. For example, the smooth regions of an image can be well reconstructed with much sparser samples, comparing to the texture-rich regions which de- serve denser sampling. In constrast, in existing practices, even for the state-of-the-art method [26] where the sampling is enforced to be locally non-uniform, it is still almost uni- form from the global perspective, which causes metadata redundancy and limits the reconstruction performance. In addition, very recent works [26, 30] sample in a fixed sam- pling space, i.e., raw image space, with a fixed bit depth of sampled pixels, leading to limited representation ability and precision redundancy. To address the above issues, instead of adopting a pre- defined sampling strategy or sampling loss, e.g., super-pixel loss [37], we propose a novel end-to-end learned raw im- age reconstruction framework based on encoded latent fea- tures. Specifically, the latent features are obtained by mini- mizing the reconstruction loss and its bitstream cost simul- taneously. To further improve the rate-distortion perfor- mance, we propose an sRGB-guided context model based on a learnable order prediction network. Different from the commonly used auto-regressive models [9, 24] which en-code/decode the latent features pixel-by-pixel in a sequen- tial way, the proposed sRGB-guided context requires much fewer steps (reduce by more than 106-fold) with the aid of a learned order mask, which makes the computational cost feasible while maintaining comparable performance. Fig. 1 compares the proposed raw image reconstruction method with the previous strategies [9, 24]. Our contributions are summarized as follows, 1. We propose the first end-to-end deep encoding frame- work for raw image reconstruction, by fully optimizing the use of stored metadata. 2. A novel sRGB-guided context model is proposed by introducing two improved entropy estimation strate- gies, which leads to better reconstruction quality, smaller size of metadata, and faster speed. 3. We evaluate our method over popular raw image datasets. The experimental results demonstrate that we can achieve better reconstruction quality with less metadata required comparing with SOTA methods.
Xiao_Level-S2fM_Structure_From_Motion_on_Neural_Level_Set_of_Implicit_CVPR_2023	Abstract This paper presents a neural incremental Structure- from-Motion (SfM) approach, Level-S2fM, which estimates the camera poses and scene geometry from a set of uncali- brated images by learning coordinate MLPs for the implicit surfaces and the radiance fields from the established key- point correspondences. Our novel formulation poses some new challenges due to inevitable two-view and few-view configurations in the incremental SfM pipeline, which com- plicates the optimization of coordinate MLPs for volumetric neural rendering with unknown camera poses. Neverthe- less, we demonstrate that the strong inductive basis convey- ing in the 2D correspondences is promising to tackle those challenges by exploiting the relationship between the ray sampling schemes. Based on this, we revisit the pipeline of incremental SfM and renew the key components, includ- ing two-view geometry initialization, the camera poses reg- istration, the 3D points triangulation, and Bundle Adjust- ment, with a fresh perspective based on neural implicit sur- faces. By unifying the scene geometry in small MLP net- works through coordinate MLPs, our Level-S2fM treats the zero-level set of the implicit surface as an informative top- down regularization to manage the reconstructed 3D points, reject the outliers in correspondences via querying SDF , and refine the estimated geometries by NBA (Neural BA). Not only does our Level-S2fM lead to promising results on camera pose estimation and scene geometry reconstruction, but it also shows a promising way for neural implicit ren- dering without knowing camera extrinsic beforehand.	1. Introduction Structure-from-Motion (SfM) is a fundamental 3D vi- sion problem that aims at reconstructing 3D scenes and estimating the camera motions from a set of uncalibrated images. As a long-standing problem, there have been a tremendous of studies that are mostly established on the keypoint correspondences across viewpoints and the theo- retical findings of Multi-View Geometry (MVG) [11], and *Corresponding author Neural Network Parametric Level Sets𝛷(x,𝜃) …Incremental Reconstruction on Neural Level SetsZero-Level SetDiscrete Points with Feature TracksDriving the Optimization SIFT Matches+RANSAC 3D Outliers from Mismatches CausesRejection by SDF ProjectionFigure 1. SfM calculations on neural level sets. We learn to do geometric calculations including Triangulation, PnP, and Bun- dle Adjustment above neural level sets, which easily help to re- ject the outliers in the matches especially in the texture repeated scenes. Also, due to the continuous surface priors of neural level sets, we achieve better pose estimation accuracy and our recon- structed points are sticking on the surface which are painted with color in the figure. While, there are a lot of outlier 3d points re- constructed by COLMAP [32] which are painted by black. have formed three representative pipelines of Incremental SfM [32], Global SfM [4, 43], and Hybrid SfM [5]. In this paper, we focus on the incremental pipeline of SfM and we will use SfM to refer to the incremental SfM. Given an unordered image set, an SfM system initializes the computation by a pair of images that are with well- conditioned keypoint correspondences to yield an initial set of feature tracks, then incrementally adds new views one by one to estimate the camera pose from the 2D-3D point correspondences and update the feature tracks with new matches. Because the feature tracks are generated by group- ing the putative 2D correspondences across viewpoints in bottom-up manners, they would be ineffective or inaccurate to represent holistic information of scenes. Accordingly, 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17205 Bundle Adjustment (BA) is necessary to jointly optimize the camera poses and 3D points in a top-down manner. The success of BA indicates that a global perspective is vital for accurate 3D reconstruction, however, their input feature tracks are the bottom-up cues without enough holistic con- straints for the 3D scenes. To this end, we study to integrate the top-down information into the SfM system by propos- ing a novel Level-S2fM. Fig. 1 illustrates a representative case for the classic SfM systems that generate more flying 3D scene points, which can be addressed by our method. Our Level-S2fM is inspired by the recently-emerged neural implicit surface that could manage all 3D scene points as the zero-level set of the signed distance function (SDF). Because the neural implicit surfaces can be param- eterized by Multi-Layer Perceptrons (MLPs), it could be viewed as a kind of top-down information of 3D scenes and is of great potential for accurate 3D reconstruction. How- ever, because both the 3D scene and camera poses are to be determined, it poses a challenging problem: How can we optimize a neural SDF (or other neu- ral fields such as NeRF) from only a set of uncal- ibrated images without any 3D information? Most recently, the above problem was partially answered in BARF [18] and NeRF- - [42] that relaxed the requirement of optimizing Neural Radiance Fields [24] without know- ing accurate camera poses, but they can only handle the un- known pose configurations in small-scale face-forwarding scenes. Moreover, when we confine the problem in the in- cremental SfM pipelines, it would be more challenging as we need to optimize the neural fields with only two over- lapped images at the initialization stage. To this end, we found that the optimization of neural SDF can be accom- plished by the 2D matches at the initialization stage, and facilitate the management of feature tracks by querying the 3D points and tracing the 2D keypoints in a holistic way. As shown in Fig. 1, we define a neural network that pa- rameterizes an SDF as the unified representation for the underdetermined 3D scene and accomplishes the computa- tions of PnP for camera pose intersection, the 3D points tri- angulation as well as the geometry refinement on the param- eterized SDF. In the initialization stage with a pair of over- lapped images, Level-S2fM uses the differentiable sphere tracing algorithm [19] to attain the corresponding 3d points of the keypoints and calculate the reprojection error to drive the joint optimization. For the traced 3d points with small SDF values and 2D reprojection errors for its feature track, they are added into a dynamic point set and take the point set with feature tracks as the Lagrangian representation for the level sets. Because the pose estimation and the scene points reconstruction are sequentially estimated, the estima- tion errors will be accumulated. To this end, we present an NBA ( i.e., Neural Bundle Adjustment) that plays a similarrole as in Bundle Adjustment, but it optimizes the implicit surface and camera poses from the explicit flow of points by the energy function of the reprojection errors, which can be viewed as an evolutionary step between Lagrangian and Eulerian representations as discussed in [23]. In the experiments, we evaluate our Level-S2fM on a va- riety of scenes from the BlendedMVS [45], DTU [14], and ETH3D [34] datasets. On the BlendedMVS dataset, our proposed Level-S2fM clearly outperforms the state-of-the- art COLMAP [32] by significant margins. On the DTU and ETH3D datasets [14, 34], our method also obtains on-par performance with COLMAP for both camera pose estima- tion and dense surface reconstruction, which are all com- puted in one stage. The contributions of this paper are in two folds: • We present a novel neural SfM approach Level-S2fM, which formulates to optimize the coordinate MLP net- works for implicit surface and radiance field and esti- mate the camera poses and scene geometry. To the best of our knowledge, our Level-S2fM is the first implicit neural SfM solution on the zero-level set of surfaces. • From the perspective of neural implicit fields learning, we show that the challenging problems of two-view and few-view optimization of neural implicit fields can be addressed by exploiting the inductive biases con- veyed in the 2D correspondences. Besides, our method presents a promising way for neural implicit rendering without knowing camera extrinsics beforehand.
Xu_High-Fidelity_Generalized_Emotional_Talking_Face_Generation_With_Multi-Modal_Emotion_Space_CVPR_2023	Abstract Recently, emotional talking face generation has received considerable attention. However, existing methods only adopt one-hot coding, image, or audio as emotion condi- tions, thus lacking flexible control in practical applications and failing to handle unseen emotion styles due to limited semantics. They either ignore the one-shot setting or the quality of generated faces. In this paper, we propose a more flexible and generalized framework. Specifically, we supple- ment the emotion style in text prompts and use an Aligned Multi-modal Emotion encoder to embed the text, image, and audio emotion modality into a unified space, which inher- its rich semantic prior from CLIP . Consequently, effective multi-modal emotion space learning helps our method sup- port arbitrary emotion modality during testing and could generalize to unseen emotion styles. Besides, an Emotion- aware Audio-to-3DMM Convertor is proposed to connect the emotion condition and the audio sequence to struc- tural representation. A followed style-based High-fidelity Emotional Face generator is designed to generate arbitrary high-resolution realistic identities. Our texture generator hierarchically learns flow fields and animated faces in a residual manner. Extensive experiments demonstrate the flexibility and generalization of our method in emotion con- trol and the effectiveness of high-quality face synthesis.	1. Introduction Talking face generation [13,38,46,58] is the task of driv- ing a static portrait with given audio. Recently, many works have tried to solve the challenges of maintaining lip move- ments synchronized with input speech contents and syn- thesizing natural facial motion simultaneously. However, most researchers ignore a more challenging task, emotional *Corresponding authorsaudio-driven talking face generation, which is critical for creating vivid talking faces. Some works have achieved significant progress in solv- ing the above task conditioned on emotion embedding. However, there are three continuously critical issues: 1) How to explore a more semantic emotion embedding to achieve better generalization for unseen emotions . Early efforts [41,47,55] adopt the one-hot vector to indicate emo- tion category, which could only cover the pre-defined la- bel and lacks semantic cues. Subsequently, EVP [19] dis- entangles emotion embedding from the audio, while GC- A VT [23] and EAMM [18] drive emotion by visual im- ages. However, tailored audio- and image-based emotion encoders show limited semantics and also struggle to handle unseen emotion styles. 2) Could we construct multi-modal emotion sources into a unified feature space to allow a more flexible and user-friendly emotion control. Existing meth- ods only support one specific modality as the emotion con- dition, while the desired modality is usually not available in practical applications. 3) How to design a high-resolution identity-generalized generator . Early works [19, 41, 47] are in identity-specific design, while recent works [18,23] have started to enable one-shot emotional talking face genera- tion. However, as shown in Fig. 1(a), GC-A VT and EAMM fail to produce high-resolution faces due to the inevitable information loss in face embedding and the challenge of es- timating accurate high-resolution flow fields, respectively. To address the aforementioned challenges, we first sup- plement the emotion styles in the text prompt inspired by the zero-shot CLIP-guided image manipulation [29,39,43], which could inherit rich semantic knowledge and conve- nient interaction after being encoded. As shown in Fig. 1(b), unseen emotions, e.g.,Satisfied , could be flexibly speci- fied using the text description and precisely reflected on the source face. Furthermore, to achieve alignment among multi-modal emotion features, we introduce an Aligned Multi-modal Emotion (AME) encoder to unify the text, im- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6609 Text:Image: Source Audio Sequence Audio: ‘Satisfied’ Emotion HappySurprisedID-1 GC-AVT (CVPR22) Single-ModalIdentity - generalized Facial embedding256 EAMM (TOG22) Single-ModalIdentity - generalized Flow Facial structure256 Ours Multi-Modal inCLIPDomainIdentity - generalizedFlow…… Facial structure512 Emotion Source ID-2Emotion Input Source Input Output Face （a）（b）Figure 1. (a) An illustrative comparison of GC-A VT [23], EAMM [18], and our approach. First, our method supports multi-modal emotion cues as input. As shown in (b), given a source face, an audio sequence, and diverse emotion conditions, our results fulfill synchronized lip movements with the speech content and emotional face with the desired style. Besides, benefiting from the effective multi-modal emotion space and rich semantics of CLIP, our method could generalize to unseen style marked in Red. Second, the hierarchical style-based generator with coarse-to-fine facial deformation learning helps us generalize to unseen faces in high resolution and provides more realistic details and precise emotion than GC-A VT and EAMM. Images are from the official attached results or released codes for fair comparisons. age, and audio emotion modality into the same domain, thus supporting flexible emotion control by multi-modal inputs, as depicted in Fig. 1(b). In particular, the fixed CLIP text and image encoders are leveraged to extract their embed- ding and a learned CLIP audio encoder guided by several losses to find the proper emotion representation of the given audio sequence in CLIP space. To this end, we follow the previous talking face gener- ation methods [34] that rely on intermediate structural in- formation such as 3DMM, and propose an Emotion-aware Audio-to-3DMM Convertor (EAC), to distill the rich emo- tional semantics from AME and project them to the facial structure. Specifically, we employ the Transformer [40] to capture the longer-term audio context and sufficiently learn correlated audio-emotion features for expression co- efficient prediction, which involves precise facial emotion and synchronized lip movement. Notably, a learned inten- sity token is extended to control the emotion intensity con- tinuously. Furthermore, to generate high-resolution realis- tic faces, we propose a coarse-to-fine style-based identity- generalized model, High-fidelity Emotional Face (HEF) generator, which integrates appearance features, geometry information, and a style code within an elegant design. As shown in Fig. 1(a), unlike the EAMM that predicts the flow field at a single resolution by an isolated process, we hier- archically perform the flow estimation in a residual manner and incorporate it with texture refinement for efficiency. In summary, we make the following three contributions: • We propose a novel AME that provides a unified multi- modal semantic-rich emotion space, allowing flexibleemotion control and unseen emotion generalization, which is the first attempt in this field. • We propose a novel HEF to hierarchically learn the facial deformation by sufficiently modeling the inter- action among emotion, source appearance, and drive geometry for the high-resolution one-shot generation. • Abundant experiments are conducted to demonstrate the superiority of our method for flexible and gener- alized emotion control, and high-resolution one-shot talking face animation over SOTA methods.
Wu_DropMAE_Masked_Autoencoders_With_Spatial-Attention_Dropout_for_Tracking_Tasks_CVPR_2023	Abstract In this paper, we study masked autoencoder (MAE) pre- training on videos for matching-based downstream tasks, including visual object tracking (VOT) and video object seg- mentation (VOS). A simple extension of MAE is to randomly mask out frame patches in videos and reconstruct the frame pixels. However, we find that this simple baseline heav- ily relies on spatial cues while ignoring temporal relations for frame reconstruction, thus leading to sub-optimal tem- poral matching representations for VOT and VOS. To al- leviate this problem, we propose DropMAE, which adap- tively performs spatial-attention dropout in the frame re- construction to facilitate temporal correspondence learning in videos. We show that our DropMAE is a strong and effi- cient temporal matching learner, which achieves better fine- tuning results on matching-based tasks than the ImageNet- based MAE with 2×faster pre-training speed. Moreover, we also find that motion diversity in pre-training videos is more important than scene diversity for improving the performance on VOT and VOS. Our pre-trained DropMAE model can be directly loaded in existing ViT-based track- ers for fine-tuning without further modifications. Notably, DropMAE sets new state-of-the-art performance on 8 out of 9 highly competitive video tracking and segmentation datasets. Our code and pre-trained models are available at https://github.com/jimmy-dq/DropMAE.git .	1. Introduction Recently, transformers have achieved enormous success in many research areas, such as natural language processing (NLP) [6, 22], computer vision [97] and audio generation [43, 73]. In NLP, masked autoencoding is commonly used to train large-scale generalizable NLP transformers contain- *Corresponding Author 60 62 64 66 68 70 72 74 76 Average Overlap (%)70727476788082848688 Success Rate (%)State-of-the-art Comparison on GOT-10k DropTrack (Ours) OSTrack SwinTrackSwinTrack MixFormerAiATrack SimTrack StarkCIA50 TransTAutoMatch SiamR-CNNOcean DiMPFigure 1. State-of-the-art comparison on GOT-10k [39] following the one-shot protocol. Our DropTrack with the proposed Drop- MAE pre-training achieves state-of-the-art performance without using complicated pipelines such as online updating. ing billions of parameters. Inspired by the great success of self-supervised learning in NLP, recent advances [36, 89] in computer vision suggest that training large-scale vision transformers may undertake a similar trajectory with NLP. The seminal work MAE [36] reconstructs the input image from a small portion of patches. The learned representa- tion in this masked autoencoder has been demonstrated to be effective in many computer vision tasks, such as image classification, object detection and semantic segmentation. In video object tracking (VOT), recently two works, SimTrack [10] and OSTrack [97], explore using an MAE pre-trained ViT model as the tracking backbone. Notably, these two trackers achieve state-of-the-art performance on existing tracking benchmarks without using complicated tracking pipelines. The key to their success is the robust pre-training weights learned by MAE on ImageNet [68]. In addition, [10, 97] also demonstrate that, for VOT, MAE un- supervised pre-training on ImageNet is more effective than supervised pre-training using class labels – this is mainly because MAE pre-training learns more fine-grained local This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14561 Input Frame PairDropMAE(Layer-4) TempMAE(Layer-6)DropMAE(Layer-6) high low TempMAE(Layer-8)DropMAE(Layer-8) TwinMAE(Layer-6) DropMAE(Layer-6) high low TwinMAE(Layer-8) DropMAE(Layer-8) Figure 2. Visualization of the attention maps of the TwinMAE baseline and our DropMAE in the reconstruction of a random masked patch, which is denoted as a red bounding box in the left input frame. TwinMAE leverages the spatial cues (within the same frame) more than temporal cues (between frames) for reconstruc- tion. Our proposed DropMAE improves the baseline by effec- tively alleviating co-adaptation between spatial cues in the recon- struction, focusing more on temporal cues, thus achieving better learning of temporal correspondences for VOT and VOS. structures that are useful for accurate target localization re- quired for VOT, whereas supervised training learns high- level class-related features that are invariant over appear- ance changes. Despite the promising performance achieved by [10,97], the MAE pre-training on ImageNet could still be sub-optimal for the tracking task due to the natural gap be- tween images and videos, i.e., no prior temporal correspon- dence information can be learned in static images. How- ever, previous tracking methods [3, 42, 83] have shown that temporal correspondence learning is the key in developing a robust and discriminative tracker. Thus there is an oppor- tunity to further develop the MAE framework specifically for matching-base video tasks, such as VOT and VOS. One simple way to extend MAE to videos is to ran- domly mask out frame patches in a video clip (i.e., video frame pairs) and then reconstruct the video clip. We de- note this simple baseline as twin MAE (TwinMAE). Given a masked patch query, as illustrated in Figs. 2 & 4, we find that TwinMAE heavily relies on the spatially neigh- bouring patches within the same frame to reconstruct the masked patch, which implies a heavy co-adaptation of spa- tial cues (within-frame tokens) for reconstruction and may cause learning of sub-optimal temporal representations for matching-based downstream tasks like video object track- ing and segmentation. To address this issue with the TwinMAE baseline, we propose DropMAE specifically designed for pre-training a masked autoencoder for matching-based video downstream tasks (e.g., VOT and VOS). Our DropMAE adaptively per- forms spatial-attention dropout to break up co-adaptation between spatial cues (within-frame tokens) during the frame reconstruction, thus encouraging more temporal interac- tions and facilitating temporal correspondence learning inthe pre-training stage. Interestingly, we obtain several im- portant findings with DropMAE: 1) DropMAE is an effec- tive and efficient temporal correspondence learner, which achieves better fine-tuning results on matching-based tasks than the ImageNet-based MAE with 2×faster pre-training speed. 2) Motion diversity in pre-training videos is more important than scene diversity for improving the perfor- mance on VOT and VOS. We conduct downstream task evaluation on 9 competi- tive VOT and VOS benchmarks, achieving state-of-the-art performance on these benchmarks. In particular, our track- ers with DropMAE pre-training obtain 75.9% AO on GOT- 10k, 52.7% AUC on LaSOT ext, 56.9% AUC on TNL2K and 92.1%/83.0% J&Fscores on DA VIS-16/17, w/o us- ing complicated online updating or memory mechanisms. In summary, the main contributions of our work are: • To the best of our knowledge, we are the first to investi- gate masked autoencoder video pre-training for tempo- ral matching-based downstream tasks. Specifically, we explore various video data sources for pre-training and build a TwinMAE baseline to study its effectiveness on temporal matching tasks. Since none exists, we further build a ViT-based VOS baseline for fine-tuning. • We propose DropMAE, which adaptively performs spatial-attention dropout in the frame reconstruction to facilitate effective temporal correspondence learn- ing in videos. • Our trackers with DropMAE pre-training sets new state-of-the-art performance on 8 out of 9 highly com- petitive video tracking and segmentation benchmarks without complicated tracking pipelines.
Wang_InternImage_Exploring_Large-Scale_Vision_Foundation_Models_With_Deformable_Convolutions_CVPR_2023	Abstract Compared to the great progress of large-scale vision transformers (ViTs) in recent years, large-scale models based on convolutional neural networks (CNNs) are still in an early state. This work presents a new large-scale CNN-based foundation model, termed InternImage, which can obtain the gain from increasing parameters and train- ing data like ViTs. Different from the recent CNNs that focus on large dense kernels, InternImage takes deformable con- volution as the core operator, so that our model not only has the large effective receptive field required for down- stream tasks such as detection and segmentation, but also has the adaptive spatial aggregation conditioned by input and task information. As a result, the proposed InternIm- age reduces the strict inductive bias of traditional CNNs and makes it possible to learn stronger and more robust patterns with large-scale parameters from massive data like ViTs. The effectiveness of our model is proven on challeng- ing benchmarks including ImageNet, COCO, and ADE20K. It is worth mentioning that InternImage-H achieved a new record 65.4 mAP on COCO test-dev and 62.9 mIoU on ADE20K, outperforming current leading CNNs and ViTs.	1. Introduction With the remarkable success of transformers in large- scale language models [3–8], vision transformers (ViTs) [2, 9–15] have also swept the computer vision field and are becoming the primary choice for the research and prac- tice of large-scale vision foundation models. Some pio- neers [16–20] have made attempts to extend ViTs to very large models with over a billion parameters, beating convo- lutional neural networks (CNNs) and significantly pushing the performance bound for a wide range of computer vision tasks, including basic classification, detection, and segmen- * equal contribution, Bcorresponding author (qiaoyu@pjlab.org.cn) (b) localattention✗long-range dependence✓adaptive spatialaggregation✓computation/memory efficient (d)dynamic sparse kernel (ours)✓long-range dependence✓adaptive spatialaggregation✓computation/memory efficient(a) globalattention✓long-range dependence✓adaptive spatialaggregation✗computation/memory efficient (c) large kernel✓long-range dependence✗adaptive spatialaggregation✓computation/memory efficientresponsepixelswithfixedweightsquerypixelsresponsepixelswithadaptiveweightsFigure 1. Comparisons of different core operators. (a) shows the global aggregation of multi-head self-attention (MHSA) [1], whose computational and memory costs are expensive in down- stream tasks that require high-resolution inputs. (b) limits the range of MHSA into a local window [2] to reduce the cost. (c) is a depth-wise convolution with very large kernels to model long- range dependencies. (d) is a deformable convolution, which shares similar favorable properties with MHSA and is efficient enough for large-scale models. We start from it to build a large-scale CNN. tation. While these results suggest that CNNs are inferior to ViTs in the era of massive parameters and data, we ar- gue that CNN-based foundation models can also achieve comparable or even better performance than ViTs when equipped with similar operator-/architecture-level designs, scaling-up parameters, and massive data . To bridge the gap between CNNs and ViTs, we first summarize their differences from two aspects: (1) From the operator level [9, 21, 22], the multi-head self-attention This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14408 (MHSA) of ViTs has long-range dependencies and adap- tive spatial aggregation (see Fig. 1(a)). Benefiting from the flexible MHSA, ViTs can learn more powerful and robust representations than CNNs from massive data. (2) From the architecture view [9, 22, 23], besides MHSA, ViTs con- tain a series of advanced components that are not included in standard CNNs, such as Layer Normalization (LN) [24], feed-forward network (FFN) [1], GELU [25], etc. Although recent works [21,22] have made meaningful attempts to in- troduce long-range dependencies into CNNs by using dense convolutions with very large kernels ( e.g., 31×31) as shown in Fig. 1 (c), there is still a considerable gap with the state- of-the-art large-scale ViTs [16, 18–20, 26] in terms of per- formance and model scale. In this work, we concentrate on designing a CNN-based foundation model that can efficiently extend to large-scale parameters and data. Specifically, we start with a flexible convolution variant—deformable convolution (DCN) [27, 28]. By combining it with a series of tailored block- level and architecture-level designs similar to transformers, we design a brand-new convolutional backbone network, termed InternImage . As shown in Fig. 1, different from recently improved CNNs with very large kernels such as 31×31 [22], the core operator of InternImage is a dynamic sparse convolution with a common window size of 3 ×3, (1) whose sampling offsets are flexible to dynamically learn ap- propriate receptive fields (can be long- or short-range) from given data; (2) the sampling offsets and modulation scalars are adaptively adjusted according to the input data, which can achieve adaptive spatial aggregation like ViTs, reduc- ing the over-inductive bias of regular convolutions; and (3) the convolution window is a common 3 ×3, avoiding the optimization problems and expensive costs caused by large dense kernels [22, 29]. With the aforementioned designs, the proposed Intern- Image can efficiently scale to large parameter sizes and learn stronger representations from large-scale training data, achieving comparable or even better performance to large-scale ViTs [2, 11, 19] on a wide range of vision tasks. In summary, our main contributions are as follows: (1) We present a new large-scale CNN-based founda- tion model—InternImage. To our best knowledge, it is the first CNN that effectively scales to over 1 billion parameters and 400 million training images and achieves comparable or even better performance than state-of-the-art ViTs, showing that convolutional models are also a worth-exploring direc- tion for large-scale model research. (2) We successfully scale CNNs to large-scale settings by introducing long-range dependencies and adaptive spa- tial aggregation using an improved 3 ×3 DCN operator, and explore the tailored basic block, stacking rules, and scaling strategies centered on the operator. These designs make ef- fective use of the operator, enabling our models to obtain 65.463.164.263.762.4 454749515355575961636567 00.511.522.533.5COCO box AP (%)#parameter (B) InternImage-H (test-dev) SwinV2 (test-dev) FD-SwinV2-G (test-dev) BEiT-3 (test-dev) Florence-CoSwin-H (test-dev) InternImage (val2017) Swin (val2017) ConvNeXt (val2017)Figure 2. Performance comparison on COCO of different backbones. The proposed InternImage-H achieves a new record 65.4 box AP on COCO test-dev, significantly outperforming state- of-the-art CNNs and large-scale ViTs. the gains from large-scale parameters and data. (3) We evaluate the proposed model on representative vision tasks including image classification, object detec- tion, instance and semantic segmentation, and compared it with state-of-the-art CNNs and large-scale ViTs by scal- ing the model size ranging from 30 million to 1 billion, the data ranging from 1 million to 400 million. Specifi- cally, our model with different parameter sizes can consis- tently outperform prior arts on ImageNet [30]. InternImage- B achieves 84.9% top-1 accuracy trained only on the ImageNet-1K dataset, outperforming CNN-based counter- parts [21, 22] by at least 1.1 points. With large-scale pa- rameters ( i.e., 1 billion) and training data ( i.e., 427 million), the top-1 accuracy of InternImage-H is further boosted to 89.6%, which is close to well-engineering ViTs [2, 19] and hybrid-ViTs [20]. In addition, on COCO [31], a challeng- ing downstream benchmark, our best model InternImage-H achieves state-of-the-art 65.4% box mAP with 2.18 billion parameters, 2.3 points higher than SwinV2-G [16] (65.4 vs. 63.1) with 27% fewer parameters as shown in Fig. 2.
Wang_DSVT_Dynamic_Sparse_Voxel_Transformer_With_Rotated_Sets_CVPR_2023	Abstract Designing an efficient yet deployment-friendly 3D back- bone to handle sparse point clouds is a fundamental problem in 3D perception. Compared with the customized sparse convolution, the attention mechanism in Transformers is more appropriate for flexibly modeling long-range relation- ships and is easier to be deployed in real-world applications. However, due to the sparse characteristics of point clouds, it is non-trivial to apply a standard transformer on sparse points. In this paper, we present Dynamic Sparse Voxel Transformer (DSVT), a single-stride window-based voxel Transformer backbone for outdoor 3D perception. In order to efficiently process sparse points in parallel, we propose Dynamic Sparse Window Attention, which partitions a series of local regions in each window according to its sparsity and then computes the features of all regions in a fully par- allel manner. To allow the cross-set connection, we design a rotated set partitioning strategy that alternates between two partitioning configurations in consecutive self-attention layers. To support effective downsampling and better en- code geometric information, we also propose an attention- style 3D pooling module on sparse points, which is powerful and deployment-friendly without utilizing any customized CUDA operations. Our model achieves state-of-the-art per- formance with a broad range of 3D perception tasks. More importantly, DSVT can be easily deployed by TensorRT with real-time inference speed (27Hz). Code will be available at https://github.com/Haiyang-W/DSVT .	1. Introduction 3D perception is a crucial challenge in computer vision, garnering increased attention thanks to its potential appli- cations in various fields such as autonomous driving sys- *Equal contribution. †Corresponding author. CenterPoint-Voxel CenterPoint-Pillar 0 5 10 15 20 25 3073 71 69 67 65 63 61 57SecondPointPillarsSSTPV-RCNN Part-A2PV-RCNN++Performance(mAPH/L2) Scene per second (Hz)VoxSetDVST-Pillar-TS (ours) DVST-Pillar (ours)DVST-Pillar-RT (ours)Figure 1. Detection performance (mAPH/L2) vs speed (Hz) of different methods on Waymo [36] validation set. All the speeds are evaluated on an NVIDIA A100 GPU with AMD EPYC 7513 CPU. tems [2, 41] and modern robotics [44, 53]. In this paper, we propose DSVT, a general-purpose and deployment-friendly Transformer-only 3D backbone that can be easily applied to various 3D perception tasks for point clouds processing. Unlike the well-studied 2D community where the input image is in a densely packed array, 3D point clouds are sparsely and irregularly distributed in continuous space due to the nature of 3D sensors, which makes it challenging to directly apply techniques used for traditional regular data. To support sparse feature learning from raw point clouds, pre- vious methods mainly apply customized sparse operations, such as PointNet++ [27, 28] and sparse convolution [12, 13]. PointNet based methods [27, 28, 39] use point-wise MLPs with the ball-query and max-pooling operators to extract features. Sparse convolution based methods [6, 12, 13] first convert point clouds to regular grids and handle the sparse volumes efficiently. Though impressive, they suffer from either the intensive computation of sampling and group- ing [28] or the limited representation capacity due to sub- manifold dilation [13]. More importantly, these specifically- designed operations generally can not be implemented with well-optimized deep learning tools ( e.g., PyTorch and Tensor- Flow) and require writing customized CUDA codes, which This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13520 greatly limits their deployment in real-world applications. Witnessing the success of transformer [40] in the 2D do- main, numerous attention-based 3D vision methods have been investigated in point cloud processing. However, be- cause of the sparse characteristic of 3D points, the number of non-empty voxels in each local region can vary significantly, which makes directly applying a standard Transformer non- trivial. To efficiently process the attention on sparse data, many approaches rebalance the token number by random sampling [25,49] or group local regions with similar number of tokens together [10, 37]. These methods are either insep- arable from superfluous computations ( e.g., dummy token padding and non-parallel attention) or noncompetitive per- formance ( e.g., token random dropping). Alternatively, some approaches [15, 24] try to solve these problems by writing customized CUDA operations, which require heavy opti- mization to be deployed on modern devices. Hence, building an efficient and deployment-friendly 3D transformer back- bone is the main challenge we aim to address in this paper. In this paper, we seek to expand the applicability of Trans- former such that it can serve as a powerful backbone for out- door 3D perception, as it does in 2D vision. Our backbone is efficient yet deployment-friendly without any customized CUDA operations. To achieve this goal, we present two major modules, one is the dynamic sparse window attention to support efficient parallel computation of local windows with diverse sparsity, and the other one is a novel learnable 3D pooling operation to downsample the feature map and better encode geometric information. Specifically, as illustrated in Figure 2, given the sparse voxelized representations and window partition, we first di- vide each window’s sparse voxels into some non-overlapped subsets, where each subset is guaranteed to have the same number of voxels for parallel computation. The partitioning configuration of these subsets will be changed in consecu- tive self-attention layers based on the rotating partition axis between the X-Axis and Y-Axis. It bridges the subsets of preceding layers for intra-window fusion, providing connec- tions among them that significantly enhance modeling power (see Table 6). To better process the inter-window fusion and encode multi-scale information, we propose the hybrid window partition, which alternates its window shape in suc- cessive transformer blocks. It leads to a drop in computation cost with even better performance (see Table 7). With the above designs, we process all regions in a fully parallel man- ner by calculating them in the same batch. This strategy is efficient in regards to real-world latency: i) all the windows are processed in parallel, which is nearly independent of the voxel distribution, ii) using self-attention without key dupli- cation, which facilitates memory access in hardware. Our experiments show that the dynamic sparse window attention approach has much lower latency than previous bucketing- based strategies [10, 37] or vanilla window attention [20], Voxel in Set 1 Voxel in Set 2Voxel in Set 3 Voxel in Set 4Local WindowX-Axis Partition Y-Axis Partition X YFigure 2. A demonstration of dynamic sparse window attention in our DSVT block. In the X-Axis DSVT layer, the sparse voxels will be split into a series of window-bounded and size-equivalent subsets in X-Axis main order, and self-attention is computed within each set. In the next layer, the set partition is switched to Y-Axis, providing connections among the previous sets. yet is similar in modeling power (see Table 5). Secondly, we present a powerful yet deployed-friendly 3D sparse pooling operation to efficiently process downsam- pling and encode better geometric representation. To tackle the sparse characteristic of 3D point clouds, previous meth- ods adopt some custom operations, e.g., customized scatter function [11] or strided sparse convolution to generate down- sampled feature volumes [46, 50]. The requirement of heavy optimization for efficient deployment limits their real-world applications. More importantly, we empirically find that inserting some linear or max-pooling layers between our transformer blocks also harms the network convergence and the encoding of geometric information. To address the above limitations, we first convert the sparse downsampling region into dense and process an attention-style 3D pooling oper- ation to automatically aggregate the local spatial features. Our 3D pooling module is powerful and deployed-friendly without any self-designed CUDA operations, and the perfor- mance gains (see Table 8) demonstrate its effectiveness. In a nutshell, our contributions are four-fold: 1) We pro- pose Dynamic Sparse Window Attention, a novel window- based attention strategy for efficiently handling sparse 3D voxels in parallel. 2) We present a learnable 3D pooling operation, which can effectively downsample sparse voxels and encode geometric information better. 3) Based on the above key designs, we introduce an efficient yet deployment- friendly transformer 3D backbone without any customized CUDA operations. It can be easily accelerated by NVIDIA TensorRT to achieve real-time inference speed ( 27Hz ), as shown in Figure 1. 4) Our approach outperforms previous state-of-the-art methods on the large-scale Waymo [35] and nuScenes [3] datasets across various 3D perception tasks.
Wang_Semantic_Scene_Completion_With_Cleaner_Self_CVPR_2023	Abstract Semantic Scene Completion (SSC) transforms an image of single-view depth and/or RGB 2D pixels into 3D vox- els, each of whose semantic labels are predicted. SSC is a well-known ill-posed problem as the prediction model has to “imagine” what is behind the visible surface, which is usually represented by Truncated Signed Distance Func- tion (TSDF). Due to the sensory imperfection of the depth camera, most existing methods based on the noisy TSDF estimated from depth values suffer from 1) incomplete vol- umetric predictions and 2) confused semantic labels. To this end, we use the ground-truth 3D voxels to generate a perfect visible surface, called TSDF-CAD, and then train a “cleaner” SSC model. As the model is noise-free, it is expected to focus more on the “imagination” of un- seen voxels. Then, we propose to distill the intermediate “cleaner” knowledge into another model with noisy TSDF input. In particular, we use the 3D occupancy feature and the semantic relations of the “cleaner self” to supervise the counterparts of the “noisy self” to respectively address the above two incorrect predictions. Experimental results validate that our method improves the noisy counterparts with3.1%IoU and 2.2%mIoU for measuring scene com- pletion and SSC, and also achieves new state-of-the-art ac- curacy on the popular NYU dataset. The code is available at https://github.com/fereenwong/CleanerS.	1. Introduction 3D scene understanding is an important visual task for many practical applications, e.g., robotic navigation [16] and augmented reality [55], where the scene geometry and semantics are two key factors to the agent interaction with the real world [24, 64]. However, visual sensors can only perceive a partial world given their limited field of view with sensory noises [49]. Therefore, an agent is expected to leverage prior knowledge to estimate the complete geome- *Corresponding author.try and semantics from the imperfect perception. Semantic Scene Completion (SSC) is designed for such an ability to infer complete volumetric occupancy and semantic labels for a scene from a single depth and/or RGB image [49, 52]. Based on an input 2D image, the 2D →3D projection is a vital bond for mapping 2D perception to the correspond- ing 3D spatial positions, which is determined by the depth value [6]. After this, the model recovers the visible surface in 3D space, which sheds light on completing and labeling the occluded regions [31, 52], because the geometry of the visible and occluded areas is tightly intertwined. For exam- ple, you can easily infer the shapes and the semantic labels when you see a part of a “chair” or “bed”. Thus, a high- quality visible surface is crucial for the SSC task. However, due to the inherent imperfection of the depth camera, the depth information is quite noisy, what follows is an imperfect visible surface that is usually represented by Truncated Signed Distance Function (TSDF) [52]. In gen- eral, the existing depth noises can be roughly categorized into the following two basic types: 1) Zero Noise . This type of noise happens when a depth sensor cannot confirm the depth value of some local regions, it will fill these regions with zeroes [14,43]. Zero noise gen- erally occurs on object surfaces with reflection or uneven- ness [41]. Based on zero noise, the visible surface will be incomplete after the 2D-3D projection via TSDF [49], so the incomplete volumetric prediction problem may occur in the final 3D voxels. For example, as shown in the upper-half of Figure 1, for the input RGB “kitchen” image, the depth value of some parts of the “cupboard” surface (marked with the red dotted frames) (in (b)) is set to zero due to reflec- tions. Based on this, both the visible surface (in (d)) and the predicted 3D voxels (in (f)) appear incomplete in reflective regions of this “cupboard”. Our method uses the perfect vis- ible surface (in (e)) generated by the noise-free ground-truth depth value (in (c)) as intermediate supervision in training , which helps the model to estimate “cupboard” 3D voxels in inference even with the noisy depth value as input. 2) Delta Noise . This type of noise refers to the inevitable deviation of the obtained depth value due to the inherent This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 867 (1) Zero Noise Leads to Incomplete Vo lumetric Predictions (b)Noisy Depth (c)Depth -CAD (d)Noisy TSDF (a)RGB (2) Delta Noise Leads to Confused Semantic Labels (b)Noisy Depth (c) Depth -CAD (d)Noisy TSDF (a)RGB (f)Prediction with (a,d) (g) Our Prediction (f)Prediction with (a, d) (g) Our Prediction table✓ chair× furniture × (e)TSDF -CAD (e)TSDF -CAD / noisy / noise- freedepth value to visible surface /incomplete /complete visible surface /visible surface with class label× ∆𝑑𝑑×Figure 1. The existing depth noises can be roughly categorized into: 1) zero noise and 2) delta noise . By zero noise , we mean that when the depth camera cannot confirm the depth value of some local regions, it fills these regions with zeroes, leading to the problem of incomplete volumetric predictions. By delta noise , we mean the inevitable deviation ( i.e.,∆d) of the obtained depth value due to the inherent quality defects of the depth camera, which leads to the problem of confused semantic labels in the final 3D voxels. In the above blocks, the pairwise subfigures (e.g., (d) and (e)) show the cases of “with noise” and “without noise” on the left and right, respectively. quality defects of the depth camera [41], i.e., the obtained depth value does not match the true depth value. Delta noise shifts the 3D position of the visible surface, resulting in the wrong semantic labels, such that the final 3D voxels will suffer from the problem of confusing semantic labels [52]. A real delta noise case is shown in the bottom half part of Figure 1. For the input RGB “classroom” image, the depth camera mistakenly estimates the depth value of the “table” as the depth value of “furniture” (in (b)). Therefore, the vis- ible surface represented by TSDF shifts from the class of “table” (marked with blue points) to the class of “furniture” (marked with orange points in (d)). Based on this, the final estimated 3D voxels (in (f)) also mistakenly estimate the part of the “table” as the “furniture”. In comparison, when our SSC model is trained on the visible surface in (e), which is generated by the correct depth value in (c), as the interme- diate supervision, semantic labels for both the “table” and the “furniture” can be estimated correctly in (g). In practice, these two types of noise are randomly mixed together to form a more complex noise [14, 65]. To handle these two noise types, although some recent SSC attempts have been made by rendering the noise-free depth value from 3D voxel ground-truth [12, 51], they are not of practi-cal use as the 3D voxels ground-truth is still needed in infer- ence. However, they indeed validate the potential that more accurate recognition performance can be achieved using the noise-free depth value [4, 56, 66]. To the best of our knowl- edge, no prior work focuses on mitigating the noisy depth values in SSC without the use of ground-truth depth val- ues in inference. Therefore, the crux is to transfer the clean knowledge learned from ground-truth depth into the noisy- depth pipeline only during training. So, in inference, we can directly use this pipeline without the need for ground-truth. In this paper, we propose a Cleaner Self (CleanerS) framework to shield the harmful effects of the two depth noises for SSC. CleanerS consists of two networks that share the same network architecture (that is what “self” means). The only difference between these two networks is that the depth value of the teacher network is rendered from ground-truth, while the depth value of the student network is inherently noisy. Therefore, the depth value of the teacher network is cleaner than the depth value of the student net- work. In the training stage, we make the teacher network to provide intermediate supervision for learning of the stu- dent network via knowledge distillation (KD), such that the student network can disentangle the clean visible surface 868 reconstruction and occluded region completion. To pre- serve both the detailed information and the abstract seman- tics of the teacher network, we adopt both feature-based andlogit-based KD strategies. In inference, only the stu- dent network is used. Compared to the noisy self, as shown in Figure 1, CleanerS achieves more accurate performance with the help of ground-truth depth values in training but not in testing. The main contributions of this work are summarized as the following two aspects: 1) we propose a novel Clean- erS framework for SSC, which can mitigate the negative effects of the noisy depth value in training; 2) CleanerS achieves the new state-of-the-art results on the challenging NYU dataset with the input of noisy depth values.
Xu_Bias-Eliminating_Augmentation_Learning_for_Debiased_Federated_Learning_CVPR_2023	Abstract Learning models trained on biased datasets tend to ob- serve correlations between categorical and undesirable fea- tures, which result in degraded performances. Most exist- ing debiased learning models are designed for centralized machine learning, which cannot be directly applied to dis- tributed settings like federated learning (FL), which col- lects data at distinct clients with privacy preserved. To tackle the challenging task of debiased federated learn- ing, we present a novel FL framework of Bias-Eliminating Augmentation Learning ( FedBEAL ), which learns to de- ploy Bias-Eliminating Augmenters ( BEA ) for producing client-specific bias-conflicting samples at each client. Since the bias types or attributes are not known in advance, a unique learning strategy is presented to jointly train BEA with the proposed FL framework. Extensive image clas- sification experiments on datasets with various bias types confirm the effectiveness and applicability of our FedBEAL, which performs favorably against state-of-the-art debiasing and FL methods for debiased FL.	1. Introduction Deep neural networks have shown promising progress across different domains such as computer vision [14] and natural language processing [8]. Their successes are typically based on the collection of and training on data that properly describe the inherent distribution of the data of interest. However, in real-world scenarios, biased data [24] are often observed during data collection. Biased datasets [10, 22, 42] contain features that are highly cor- related to class labels in the training dataset but not suffi- ciently describing the inherent semantic meaning. Training on such biased data thus result in degraded model general- ization capability. Take Fig. 1 for example; when address- ing the cat-dog classification task, training images collected by users might contain only orange cats and black dogs. Their color attributes are strongly correlated with the image labels during training, but such attributes are not necessar- ily relevant to the classification task during inference. As Figure 1. Example of local data bias in FL. When deploying FL to train a cat-dog classifier with image datasets collected by multiple pet owners, most of the local images are obtained with their pets with specific colors. Therefore, the models trained with each local dataset are likely to establish decision rules on biased attributes ( e.g., fur color), which prevents the aggregated model from learning proper representation for classification. pointed out in [10, 42], deep neural networks trained with such biased data are more likely to make decisions based onbias attributes instead of semantic attributes. As a re- sult, during inference, performances of the learned models would dramatically drop when observing bias-conflicting samples ( i.e., data containing semantic and bias attributes that are rarely correlated in the training set). To tackle the data bias problem, several works have been proposed to remove or alleviate data bias when training deep learning models [6, 11, 18, 24, 27, 32, 36, 40]. For ex- ample, Nam et al. [36] train an intentionally biased auxil- iary model while enforcing the main model to go against the prejudice of the biased network. Lee et al. [27] utilize the aforementioned biased model to synthesize diverse bias- conflicting hidden features for learning debiased represen- tations. Nevertheless, the above techniques are designed for centralized datasets. When performing distributed training of learning models, such methods might fail to generalize. For distributed learning, federated learning (FL) [35] This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20442 particularly considers data collection and training con- ducted at each client, with data privacy needing to be pre- served. When considering privately distributed datasets, real-world FL applications are more likely to suffer data heterogeneity issues [20, 28, 51], i.e., data collected by different clients are not independent and identically dis- tributed (IID). Recently, several works [19,21,29–31,34,47] propose to alleviate performance degradation caused by data heterogeneity. However, existing methods typically consider data heterogeneity in terms of label distribution skew [21, 29, 30, 34, 47] or domain discrepancy [19, 31] among clients. These FL methods are not designed to tackle potential data bias across different clients, leaving the debi- ased FL a challenging task to tackle. To mitigate the local bias in Federated learning, we pro- pose a novel FL scheme of Bias-Eliminating Augmentation Learning ( FedBEAL ). In FedBEAL, we learn a Bias- Eliminating Augmenter (BEA) for each client, with the goal of producing bias-conflicting samples. To identify and in- troduce the desirable semantic and bias attributes to the aug- mented samples, our FedBEAL uniquely adopts the global server model and each client model trained across iterations without prior knowledge of bias type or annotation. With the introduced augmenter and the produced bias-conflicting samples, debiased local updates can be performed at each client, followed by simple aggregation of such models for deriving the server model. We now summarize the contributions of this work below: • To the best of our knowledge, We are among the first to tackle the problem of debiased federated learning, in which local yet distinct biases exist at the client level. • We present FedBEAL for debiased FL, which intro- duces Bias-Eliminating Augmenters (BEA) at each client with the goal of generating bias-conflicting sam- ples to eliminate local data biases. • Learning of BEA can be realized by utilizing the global server and local client models trained across iterations, which allows us to identify and embed desirable se- mantic and bias features for augmentation purposes.
Weder_Removing_Objects_From_Neural_Radiance_Fields_CVPR_2023	Abstract Neural Radiance Fields (NeRFs) are emerging as a ubiq- uitous scene representation that allows for novel view syn- thesis. Increasingly, NeRFs will be shareable with other people. Before sharing a NeRF , though, it might be desir- able to remove personal information or unsightly objects. Such removal is not easily achieved with the current NeRF editing frameworks. We propose a framework to remove objects from a NeRF representation created from an RGB- D sequence. Our NeRF inpainting method leverages re- cent work in 2D image inpainting and is guided by a user- provided mask. Our algorithm is underpinned by a confi- dence based view selection procedure. It chooses which of the individual 2D inpainted images to use in the creation of the NeRF , so that the resulting inpainted NeRF is 3D consis- tent. We show that our method for NeRF editing is effective for synthesizing plausible inpaintings in a multi-view co- herent manner, outperforming competing methods. We vali- date our approach by proposing a new and still-challenging dataset for the task of NeRF inpainting.	1. Introduction Since the initial publication of Neural Radiance Fields (NeRFs) [42], there has been an explosion of extensions to the original framework, e.g., [3, 4, 8, 12, 25, 35, 39, 42]. NeRFs are being used beyond the initial task of novel view synthesis. It is already appealing to get them into the hands of non-expert users for novel applications, e.g., for NeRF editing [80] or live capture and training [47], and these more casual use cases are driving interesting new technical issues. One of those issues is how to seamlessly remove parts of the rendered scene. Removing parts of the scene can be desirable for a variety of reasons. For example, a house scan being shared on a property selling website may need unappealing or personally identifiable objects to be re- moved [68]. Similarly, objects could be removed so they can be replaced in an augmented reality application, e.g., removing a chair from a scan to see how a new chair fits Input images NeRF Ours Novel view Novel view with the object removed Novel views Novel views without the object Ours Input images NeRF Figure 1. Removal of unsightly objects. Our method allows for objects to be plausibly removed from NeRF reconstructions, in- painting missing regions whilst preserving multi-view coherence. in the environment [51]. Removing objects might also be desirable when a NeRF is part of a traditional computer vi- sion pipeline, e.g., removing parked cars from scans that are going to be used for relocalization [44]. Some editing of NeRFs has already been explored. For example, object-centric representations disentangle labeled objects from the background, which allows editing of the trained scene with user-guided transformations [74, 77], while semantic decomposition allows selective editing and transparency for certain semantic parts of the scene [26]. However, these previous approaches only augment informa- tion from the input scan, limiting their generative capabil- ities, i.e.,the hallucination of elements that have not been observed from any view. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16528 With this work, we tackle the problem of removing ob- jects from scenes, while realistically filling the resulting holes, as shown in Fig. 1. Solving this problem requires: a) exploiting multi-view information when parts of the scene are observed in some frames but occluded in others and, b) leveraging a generative process to fill areas that are never observed. To this end, we pair the multi-view consistency of NeRFs with the generative power of 2D inpainting mod- els [69] that are trained on large scale 2D image datasets. Such 2D inpaintings are not multi-view consistent by con- struction, and may contain severe artefacts. Using these in- paintings directly causes corrupted reconstructions, so we design a new confidence-based view-selection scheme that iteratively removes inconsistent inpaintings from the opti- mization. We validate our approach on a new dataset and show that we outperform existing approaches for novel view synthesis on standard metrics of image quality, as well as producing multi-view consistent results. In summary, we make the following contributions: 1) We propose the first approach focusing on inpainting NeRFs by leveraging the power of single image inpainting. 2) We introduce a novel view-selection mechanism that au- tomatically removes inconsistent views from the optimiza- tion. 3) We present a new dataset for evaluating object re- moval and inpainting in indoor and outdoor scenes.
Wang_A_Practical_Upper_Bound_for_the_Worst-Case_Attribution_Deviations_CVPR_2023	Abstract Model attribution is a critical component of deep neural networks (DNNs) for its interpretability to complex models. Recent studies bring up attention to the security of attribu- tion methods as they are vulnerable to attribution attacks that generate similar images with dramatically different at- tributions. Existing works have been investigating empir- ically improving the robustness of DNNs against those at- tacks; however, none of them explicitly quantifies the actual deviations of attributions. In this work, for the first time, a constrained optimization problem is formulated to derive an upper bound that measures the largest dissimilarity of attributions after the samples are perturbed by any noises within a certain region while the classification results re- main the same. Based on the formulation, different prac- tical approaches are introduced to bound the attributions above using Euclidean distance and cosine similarity un- der both ℓ2andℓ∞-norm perturbations constraints. The bounds developed by our theoretical study are validated on various datasets and two different types of attacks (PGD at- tack and IFIA attribution attack). Over 10 million attacks in the experiments indicate that the proposed upper bounds effectively quantify the robustness of models based on the worst-case attribution dissimilarities.	1. Introduction Attribution methods play an important role in deep learn- ing applications as one of the subareas of explainable AI. Practitioners use attribution methods to measure the rela- tive importance among different features and to understand the impacts of features contributing to the model outputs. They have been widely used in a number of critical real- world applications, such as risk management [2], medical imaging [24, 29] and drug discovery [13]. In particular, attributions are supposed to be secure and resistant to ex- ternal manipulation such that proper explanations can be applied to safety-sensitive applications. Regulations arealso deployed in countries to enforce the interpretability of deep learning models for a ‘right to explain’ [10]. Al- though attribution methods have been extensively studied [18,25,28,31,36,39], recent works reveal that they are vul- nerable to visually imperceptible perturbations that drasti- cally alter the attributions and keep the model outputs un- changed [6, 8]. Prior works [3, 4, 12, 23, 30, 32, 33] investigate the attri- bution robustness based on empirical and statistical estima- tions over entire dataset. However, current attribution ro- bustness works are unable to evaluate how robust the model is given any arbitrary test point, perturbed or unperturbed. In this paper, we study the problem of finding the worst attribution perturbation within certain predefined regions. Specifically, given a trained model and an image sample, we propose theoretical upper bounds of the attribution de- viations from the unperturbed ones. As far as we know, this is the first attempt to provide an upper bound of attribution differences. In this paper, the general upper bound for attribution de- viation is first quantified as the maximum changes of attri- butions after the samples are perturbed while classification results remain the same. Two cases are analyzed, including with and without label constraint, which refers to the classi- fication labels being unchanged and changed, respectively, after the original samples are attacked. For each case, two mostly used perturbation constraints, ℓ2andℓ∞-norm, are considered to compute the upper bound. For ℓ2-norm con- straint, our approach is based on the first-order Taylor series of model attribution, and a tight upper bound ignoring the label constraint is computed from the singular value of the attribution gradient. ℓ∞-norm constraint is more compli- cated because the upper bound is a solution of a concave quadratic programming with box constraints, which is an NP-hard problem. Thus, two relaxation approaches are pro- posed. Moreover, a more restricted bound constrained on the unchanged label is also studied. In this study, Euclidean distance and cosine distance, which are also employed in the previous empirical studies [4, 30, 32], are used as dis- similarity functions to measure attribution difference. We This CVPR paper is the Open Access version, provided by the Co
Wei_Text-Guided_Unsupervised_Latent_Transformation_for_Multi-Attribute_Image_Manipulation_CVPR_2023	Abstract Great progress has been made in StyleGAN-based im- age editing. To associate with preset attributes, most ex- isting approaches focus on supervised learning for seman- tically meaningful latent space traversal directions, and each manipulation step is typically determined for an in- dividual attribute. To address this limitation, we propose a Text-guided Unsupervised StyleGAN Latent Transformation (TUSLT) model, which adaptively infers a single transfor- mation step in the latent space of StyleGAN to simultane- ously manipulate multiple attributes on a given input image. Speciﬁcally, we adopt a two-stage architecture for a latent mapping network to break down the transformation process into two manageable steps. Our network ﬁrst learns a di- verse set of semantic directions tailored to an input image, and later nonlinearly fuses the ones associated with the tar- get attributes to infer a residual vector. The resulting tightly interlinked two-stage architecture delivers the ﬂexibility to handle diverse attribute combinations. By leveraging the cross-modal text-image representation of CLIP , we can per- form pseudo annotations based on the semantic similarity between preset attribute text descriptions and training im- ages, and further jointly train an auxiliary attribute clas- siﬁer with the latent mapping network to provide semantic guidance. We perform extensive experiments to demonstrate that the adopted strategies contribute to the superior perfor- mance of TUSLT.	1. Introduction Visual attributes represent semantically meaningful fea- tures inherent in images, and attribute manipulation has ex- Corresponding author. Figure 1. Visually comparing TUSLT with StyleFlow (supervised) and StyleCLIP (text-driven) in precisely manipulating multiple at- tributes and preserving irrelevant attributes. perienced great improvements, due to the advent of Gen- erative Adversarial Network [13] (GAN)-based generative models, e.g. StyleGAN [21, 22] and StarGAN [7, 8]. Re- cent works [15,37,43] have discovered that the latent space of StyleGAN possesses semantic disentanglement proper- ties, enabling a variety of image editing operations via la- tent transformations. StyleGAN-based methods for image attribute manipula- tion typically involve a large number of manual annotations or well-trained attribute classiﬁers. Furthermore, the dis- covered semantic latent directions are associated with indi- vidual attributes. The editing on a target attribute is car- ried out by moving the latent code of an input image along one of the directions. For Ktarget attributes, these model- s requireKtransformation steps to handle the translation. As a result, they are not scalable to the increasing number of target attributes in multi-attribute transformation tasks. As shown in Figure 1, we test a state-of-the-art supervised This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19285 Figure 2. Overview of the proposed model, TUSLT, consisting of two learnable components: an auxiliary attribute classiﬁer Atrained on the CLIP-based labeled data, and a latent mapping network f;g.infers latent directions f (1) w;:::; (K) wgfor preset attributes, and transforms the target-related directions as indicated by mask Minto a residual vector w, to which the initial latent code is added. Precise multi-attribute transfer is allowed by such a single transformation step, and the generator Gsynthesizes a new image reﬂecting the target attributes under the guidance of Aand CLIP encoders. model, StyleFlow [2], and ﬁnd that multiple transformation steps lead to undesired deviation from the input image on irrelevant attributes. Compared to the state-of-the-art text- driven model, StyleCLIP [33], we can also achieve a better manipulation result by seeking a single latent transforma- tion step for the task. More speciﬁcally, we propose a Text-guided Unsuper- vised StyleGAN Latent Transformation (TUSLT) model that supports simultaneous manipulation on multiple at- tributes. As shown in Figure 2, the key is to jointly learn a mapping network to infer the latent transformation and an auxiliary attribute classiﬁer to assess manipulation quality. We employ the Contrastive Language-Image Pre-training (CLIP) model [34] to generate pseudo-labeled data by mea- suring the semantic similarities between attribute text de- scriptions and training images. Compared to CLIP, the jointly trained classiﬁer extracts domain-speciﬁc informa- tion to better characterize the differences among attributes. This beneﬁts the mapping network to seek more suitable transformations, such that the synthesized images reﬂec- t target attributes. Further, we adopt a two-stage architec- ture for the mapping network: the earlier stage employs a prediction subnetwork to infer a set of semantic direction- s, and the latter stage operates on the resulting directions and nonlinearly fuses the target-related ones. The inter- mediate semantic directions are associated with preset at- tributes and tailored for the input image. This design allows us to deal with a wide range of attribute combinations in a single transformation step. We perform extensive experi- ments and provide both qualitative and quantitative results in diverse multi-attribute transformation tasks, showing thesuperiority of our model over the competing methods. In summary, the main contributions of this work are giv- en as follows: (a) The existing image editing methods focus on discovering semantic latent directions associated with individual visual attributes, and a sequential manipulation process is thus needed for multi-attribute manipulation. In contrast, the proposed model infers a single step of latent s- pace walk to simultaneously manipulate multiple attributes. (b) Beneﬁting from the cross-modal text-image representa- tion of CLIP, we jointly train a latent mapping network with an auxiliary attribute classiﬁer, which leads to more precise attribute rendering without requiring additional manual an- notations. (c) Due to the two-stage nature, our latent map- ping network breaks down the challenging multi-attribute manipulation task into sub-tasks: inferring diverse seman- tic directions and integrating the target-related ones into a single transformation vector. This design gives our model interpretability and ﬂexibility in dealing with a variety of attribute combinations.
Wang_MoLo_Motion-Augmented_Long-Short_Contrastive_Learning_for_Few-Shot_Action_Recognition_CVPR_2023	Abstract Current state-of-the-art approaches for few-shot action recognition achieve promising performance by conducting frame-level matching on learned visual features. However, they generally suffer from two limitations: i) the matching procedure between local frames tends to be inaccurate due to the lack of guidance to force long-range temporal percep- tion; ii) explicit motion learning is usually ignored, leading to partial information loss. To address these issues, we de- velop a Motion-augmented Long-short Contrastive Learn- ing (MoLo) method that contains two crucial components, including a long-short contrastive objective and a motion autodecoder. Specifically, the long-short contrastive objec- tive is to endow local frame features with long-form tem- poral awareness by maximizing their agreement with the global token of videos belonging to the same class. The motion autodecoder is a lightweight architecture to recon- struct pixel motions from the differential features, which ex- plicitly embeds the network with motion dynamics. By this means, MoLo can simultaneously learn long-range tempo- ral context and motion cues for comprehensive few-shot matching. To demonstrate the effectiveness, we evaluate MoLo on five standard benchmarks, and the results show that MoLo favorably outperforms recent advanced meth- ods. The source code is available at https://github. com/alibaba-mmai-research/MoLo .	1. Introduction Recently, action recognition has achieved remarkable progress and shown broad prospects in many application fields [1, 5, 8, 37, 67]. Despite this, these successes rely heavily on large amounts of manual data annotation, which greatly limits the scalability to unseen categories due to the ∗Intern at Alibaba DAMO Academy. †Corresponding authors. Support video: “ Picking something up ” Query video is m isclassified as “Picking something up” Real label: “ Removing something, revealing something behind”Match? Match?Match?Match? Match? Match? (a) Failure case one Support video: “Pushing something from right to left” Query video is m isclassified as “Pushing something from right to left ” Real label: “ Tipping something over”Match? Match?Match?Match? Match? Match? (b) Failure case two Figure 1. Illustration of our motivation. We show that most existing metric-based local frame matching methods, such as OTAM [4], can be easily perturbed by some similar co-existing video frames due to the lack of forced global context awareness during the support-query temporal alignment process. Example videos come from the commonly used SSv2 dataset [15]. high cost of acquiring large-scale labeled samples. To alle- viate the reliance on massive data, few-shot action recogni- tion [88] is a promising direction, aiming to identify novel classes with extremely limited labeled videos. Most mainstream few-shot action recognition ap- proaches [4,21,44,74] adopt the metric-based meta-learning strategy [61] that learns to map videos into an appropriate feature space and then performs alignment metrics to pre- dict query labels. Typically, OTAM [4] leverages a deep network to extract video features and explicitly estimates an ordered temporal alignment path to match the frames of two videos. HyRSM [74] proposes to explore task-specific semantic correlations across videos and designs a bidirec- tional Mean Hausdorff Metric (Bi-MHM) to align frames. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18011 Though these works have obtained significant results, there are still two limitations: first, existing standard metric-based techniques mainly focus on local frame-level alignment and are considered limited since the essential global informa- tion is not explicitly involved. As shown in Figure 1, lo- cal frame-level metrics can be easily affected by co-existing similar video frames. We argue that it would be beneficial to achieve accurate matching if the local frame features can predict the global context in few-shot classification; sec- ond, motion dynamics are widely regarded as a vital role in the field of video understanding [6, 22, 25, 42, 50, 70], while the existing few-shot methods do not explicitly ex- plore the rich motion cues between frames for the matching procedure, resulting in a sub-optimal performance. In the literature [5, 67], traditional action recognition works intro- duce motion information by feeding optical flow or frame difference into an additional deep network, which leads to non-negligible computational overhead. Therefore, an effi- cient motion compensation method should be introduced to achieve comprehensive few-shot matching. Inspired by the above observations, we develop a motion-augmented long-short contrastive learning (MoLo) method to jointly model the global contextual information and motion dynamics. More specifically, to explicitly in- tegrate the global context into the local matching process, we apply a long-short contrastive objective to enforce frame features to predict the global context of the videos that be- long to the same class. For motion compensation, we de- sign a motion autodecoder to explicitly extract motion fea- tures between frame representations by reconstructing pixel motions, e.g., frame differences. In this way, our proposed MoLo enables efficient and comprehensive exploitation of temporal contextual dependencies and motion cues for ac- curate few-shot action recognition. Experimental results on multiple widely-used benchmarks demonstrate that our MoLo outperforms other advanced few-shot techniques and achieves state-of-the-art performance. In summary, our contributions can be summarized as fol- lows: (1) We propose a novel MoLo method for few-shot action recognition, aiming to better leverage the global con- text and motion dynamics. (2) We further design a long- short contrastive objective to reinforce local frame features to perceive comprehensive global information and a motion autodecoder to explicitly extract motion cues. (3) We con- duct extensive experiments across five widely-used bench- marks to validate the effectiveness of the proposed MoLo. The results demonstrate that MoLo significantly outper- forms baselines and achieves state-of-the-art performance.
Xing_CodeTalker_Speech-Driven_3D_Facial_Animation_With_Discrete_Motion_Prior_CVPR_2023	Abstract Speech-driven 3D facial animation has been widely stud- ied, yet there is still a gap to achieving realism and vividness due to the highly ill-posed nature and scarcity of audio- visual data. Existing works typically formulate the cross- modal mapping into a regression task, which suffers from the regression-to-mean problem leading to over-smoothed facial motions. In this paper, we propose to cast speech- driven facial animation as a code query task in a finite proxy space of the learned codebook, which effectively pro- motes the vividness of the generated motions by reducing the cross-modal mapping uncertainty. The codebook is learned by self-reconstruction over real facial motions and thus embedded with realistic facial motion priors. Over the discrete motion space, a temporal autoregressive model is employed to sequentially synthesize facial motions from the input speech signal, which guarantees lip-sync as well as plausible facial expressions. We demonstrate that our ap- proach outperforms current state-of-the-art methods both qualitatively and quantitatively. Also, a user study fur- ther justifies our superiority in perceptual quality. Code and video demo are available at https://doubiiu. github.io/projects/codetalker .	1. Introduction 3D facial animation has been an active research topic for decades, as attributed to its broad applications in virtual re- ality, film production, and games. The high correlation be- tween speech and facial gestures (especially lip movements) makes it possible to drive the facial animation with a speech signal. Early attempts are mainly made to build the complex mapping rules between phonemes and their visual counter- part, which usually have limited performance [53,63]. With the advances in deep learning, recent speech-driven facial animation techniques push forward the state-of-the-art sig- nificantly. However, it still remains challenging to generate human-like motions. *Corresponding Author.As an ill-posed problem, speech-driven facial animation generally has multiple plausible outputs for every input. Such ambiguity tends to cause over-smoothed results. Any- how, person-specific approaches [29, 49] can usually ob- tain decent facial motions because of the relatively consis- tent talking style, but have low scalability to general ap- plications. Recently, VOCA [10] extends these methods to generalize across different identities, however, they gen- erally exhibit mild or static upper face expressions. This is because VOCA formulates the speech-to-motion map- ping as a regression task, which encourages averaged mo- tions, especially in the upper face that is only weakly or even uncorrelated to the speech signal. To reduce the un- certainty, FaceFormer [16] utilizes long-term audio context through a transformer-based model and synthesizes the se- quential motions in an autoregressive manner. Although it gains important performance promotion, it still inherits the weakness of one-to-one mapping formulation and suf- fers from a lack of subtle high-frequency motions. Dif- ferently, MeshTalk [50] models a categorical latent space for facial animation that disentangles audio-correlated and audio-uncorrelated information so that both aspects could be well handled. Anyway, the employed quantization and categorical latent space representation are not well-suited for motion prior learning, rendering the training tricky and consequently hindering its performance. We get inspiration from 3D Face Morphable Model (3DMM) [35], where general facial expressions are rep- resented in a low-dimensional space. Accordingly, we propose to formulate speech-driven facial animation as a code query task in a finite proxy space of the learned dis- crete codebook prior. The codebook is learned by self- reconstruction over real facial motions using a vector- quantized autoencoder (VQ-V AE) [57], which along with the decoder stores the realistic facial motion priors. In contrast to the continuous linear space of 3DMM, com- binations of codebook items form a discrete prior space with only finite cardinality. Still, in the context of the de- coder, the code representation possesses high expressive- ness. Through mapping the speech to the finite proxy space, the uncertainty of the speech-to-motion mapping is signif- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12780 icantly attenuated and hence promotes the quality of mo- tion synthesis. Conceptually, the proxy space approximates the facial motion space, where the learned codebook items serve as discrete motion primitives. Based on the learned discrete codebook, we pro- pose a code-query-based temporal autoregressive model for speech-conditioned facial motion synthesis, called CodeTalker . Specifically, taking a speech signal as input, our model predicts the motion feature tokens in a temporal recursive manner. Then, the feature tokens are used to query the code sequence in the discrete space, followed by facial motion reconstruction. Thanks to the contextual modeling over history motions and cross-modal alignment, the pro- posed CodeTalker shows the advantages of achieving accu- rate lip motions and natural expressions. Extensive experi- ments show that the proposed CodeTalker demonstrates su- perior performance on existing datasets. Systematic studies and experiments are conducted to demonstrate the merits of our method over previous works. The contributions of our work are as follows: • We model the facial motion space with discrete prim- itives in a novel way, which offers advantages to pro- mote motion synthesis realism against cross-modal un- certainty. • We propose a discrete motion prior based temporal au- toregressive model for speech-driven facial animation, which outperforms existing state-of-the-art methods.
Wang_Generalist_Decoupling_Natural_and_Robust_Generalization_CVPR_2023	Abstract Deep neural networks obtained by standard training have been constantly plagued by adversarial examples. Although adversarial training demonstrates its capability to defend against adversarial examples, unfortunately, it leads to an inevitable drop in the natural generalization. To address the issue, we decouple the natural generalization and the robust generalization from joint training and formulate different training strategies for each one. Specifically, instead of min- imizing a global loss on the expectation over these two gen- eralization errors, we propose a bi-expert framework called Generalist where we simultaneously train base learners with task-aware strategies so that they can specialize in their own fields. The parameters of base learners are collected and combined to form a global learner at intervals during the training process. The global learner is then distributed to the base learners as initialized parameters for continued train- ing. Theoretically, we prove that the risks of Generalist will get lower once the base learners are well trained. Extensive experiments verify the applicability of Generalist to achieve high accuracy on natural examples while maintaining con- siderable robustness to adversarial ones. Code is available athttps://github.com/PKU-ML/Generalist .	1. Introduction Modern deep learning techniques have achieved remark- able success in many fields, including computer vision [14, 16], natural language processing [10, 31], and speech recognition [28, 36]. Yet, deep neural networks (DNNs) suffer a catastrophic performance degradation by human imperceptible adversarial perturbations where wrong predic- tions are made with extremely high confidence [13, 29, 34]. The vulnerability of DNNs has led to the proposal of var- ious defense approaches [3, 24, 25, 33, 40] for protecting DNNs from adversarial attacks. One of those representa- *Work was done as an internship at Peking University. Now, he is a Ph.D. student at the University of Hong Kong. †Corresponding Author: Yisen Wang (yisen.wang@pku.edu.cn) 83 84 85 86 87 88 89 Natural Accuracy32.535.037.540.042.545.047.5Robust Accuracy (AA) Madry TRADES FAT GeneralistFigure 1. Comparison with other advanced adversarial training methods. Both clean accuracy and robust accuracy (against Au- toAttack [9]) are given for a handy reference. It is noted that current adversarial training methods achieve high clean accuracy by greatly sacrificing robustness. That means it is hard to obtain sufficient robustness but maintain high clean accuracy in the joint training framework. Our Generalist attains excellent clean accuracy while staying competitively robust. The improvement of Generalist is notable since we only use the naive cross-entropy loss with negligi- ble computational overhead and even without increasing the model size. tive techniques is adversarial training (AT) [20, 21, 37, 38], which dynamically injects perturbed examples that deceive the current model but preserve the right label into the train- ing set. Adversarial training has been demonstrated to be the most effective method to improve adversarially robust generalization [2, 39]. Despite these successes, such attempts of adversarial training have found a tradeoff between natural and robust accuracy, i.e., there exists an undesirable increase in the er- ror on unperturbed images when the error on the worst-case perturbed images decreases, as illustrated in Figure 1. Prior works [30, 43] even argue that natural and robust accuracy are fundamentally at odds, which indicates that a robust clas- sifier can be achieved only when compromising the natural generalization. However, the following works found that This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20554 the tradeoff may be settled in a roundabout way, such as incorporating additional labeled/unlabeled data [1, 8, 22, 26] or relaxing the magnitude of perturbations to generate suit- able adversarial examples for better optimization [18, 44]. These works all focus on the data used for training while we propose to tackle the tradeoff problem from the perspective of the
Wang_BAD-NeRF_Bundle_Adjusted_Deblur_Neural_Radiance_Fields_CVPR_2023	Abstract Neural Radiance Fields (NeRF) have received consider- able attention recently, due to its impressive capability in photo-realistic 3D reconstruction and novel view synthesis, given a set of posed camera images. Earlier work usually assumes the input images are of good quality. However, im- age degradation (e.g. image motion blur in low-light con- ditions) can easily happen in real-world scenarios, which would further affect the rendering quality of NeRF . In this paper, we present a novel bundle adjusted deblur Neural Radiance Fields (BAD-NeRF), which can be robust to se- vere motion blurred images and inaccurate camera poses. Our approach models the physical image formation process of a motion blurred image, and jointly learns the parameters of NeRF and recovers the camera motion trajectories dur- ing exposure time. In experiments, we show that by directly modeling the real physical image formation process, BAD- NeRF achieves superior performance over prior works on both synthetic and real datasets. Code and data are avail- able at https://github.com/WU-CVGL/BAD-NeRF .	1. Introduction Acquiring accurate 3D scene geometry and appearance from a set of 2D images has been a long standing problem †Corresponding author.in computer vision. As a fundamental block for many vi- sion applications, such as novel view image synthesis and robotic navigation, great progress has been made over the last decades. Classic approaches usually represent the 3D scene explicitly, in the form of 3D point cloud [8, 52], tri- angular mesh [4, 5, 10] or volumetric grid [31, 45]. Recent advancements in implicit 3D representation by using a deep neural network, such as Neural Radiance Fields (NeRF) [27], have enabled photo-realistic 3D reconstruction and novel view image synthesis, given well posed multi-view images. NeRF takes a 5D vector (i.e. for spatial location and viewing direction of the sampled 3D point) as input and predicts its radiance and volume density via a multilayer perceptron. The corresponding pixel intensity or depth can then be computed by differentiable volume rendering [19, 25]. While many methods have been proposed to fur- ther improve NeRF’s performance, such as rendering effi- ciency [11,28], training with inaccurate poses [20] etc., lim- ited work has been proposed to address the issue of training with motion blurred images. Motion blur is one of the most common artifacts that degrades images in practical appli- cation scenarios. It usually occurs in low-light conditions where longer exposure times are necessary. Motion blurred images would bring two main challenges to existing NeRF training pipeline: a) NeRF usually assumes the rendered This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4170 image is sharp (i.e. infinitesimal exposure time), motion blurred image thus violates this assumption; b) accurate camera poses are usually required to train NeRF, however, it is difficult to obtain accurate poses from blurred images only, since each of them usually encodes information of the motion trajectory during exposure time. On the other hand, it is also challenging itself to recover accurate poses (e.g., via COLMAP [41]) from a set of motion blurred images, due to the difficulties of detecting and matching salient key- points. Combining both factors would thus further degrade NeRF’s performance if it is trained with motion blurred im- ages. In order to address those challenges, we propose to in- tegrate the real physical image formation process of a mo- tion blurred image into the training of NeRF. We also use a linear motion model in the SE(3) space to represent the camera motion trajectory within exposure time. During the training stage, both the network weights of NeRF and the camera motion trajectories are estimated jointly. In partic- ular, we represent the motion trajectory of each image with both poses at the start and end of the exposure time respec- tively. The intermediate camera poses within exposure time can be linearly interpolated in the SE(3) space. This as- sumption holds in general since the exposure time is typ- ically small. We can then follow the real physical image formation model of a motion blurred image to synthesize the blurry images. In particular, a sequence of sharp im- ages along the motion trajectory within exposure time can be rendered from NeRF. The corresponding motion blurred image can then be synthesized by averaging those virtual sharp images. Both NeRF and the camera motion trajec- tories are estimated by minimizing the difference between the synthesized blurred images and the real blurred images. We refer this modified model as BAD-NeRF, i.e. bundle adjusted deblur NeRF. We evaluate BAD-NeRF with both synthetic and real datasets. The experimental results demonstrate that BAD- NeRF achieves superior performance compared to prior state of the art works (e.g. as shown in Fig. 1), by explicitly modeling the image formation process of the motion blurred image. In summary, our contributions are as follows: • We present a photo-metric bundle adjustment formula- tion for motion blurred images under the framework of NeRF, which can be potentially integrated with other vision pipelines (e.g. a motion blur aware camera pose tracker [21]) in future. • We show how this formulation can be used to acquire high quality 3D scene representation from a set of mo- tion blurred images. • We experimentally validate that our approach is able to deblur severe motion blurred images and synthesize high quality novel view images.
Wang_Deep_Factorized_Metric_Learning_CVPR_2023	Abstract Learning a generalizable and comprehensive similarity metric to depict the semantic discrepancies between images is the foundation of many computer vision tasks. While ex- isting methods approach this goal by learning an ensemble of embeddings with diverse objectives, the backbone net- work still receives a mix of all the training signals. Differ- ently, we propose a deep factorized metric learning (DFML) method to factorize the training signal and employ different samples to train various components of the backbone net- work. We factorize the network to different sub-blocks and devise a learnable router to adaptively allocate the training samples to each sub-block with the objective to capture the most information. The metric model trained by DFML cap- ture different characteristics with different sub-blocks and constitutes a generalizable metric when using all the sub- blocks. The proposed DFML achieves state-of-the-art per- formance on all three benchmarks for deep metric learn- ing including CUB-200-2011, Cars196, and Stanford On- line Products. We also generalize DFML to the image clas- sification task on ImageNet-1K and observe consistent im- provement in accuracy/computation trade-off. Specifically, we improve the performance of ViT-B on ImageNet (+0.2% accuracy) with less computation load (-24% FLOPs).1	1. Introduction Learning good representations for images has always been the core of computer vision, yet measuring the sim- ilarity between representations after obtaining them is an equally important problem. Focusing on this, metric learn- ing aims to learn a discriminative similarity metric un- der which the interclass distances are large and the intra- class distances are small. Using a properly learned simi- larity metric can improve the performance of downstream tasks and has been employed in many applications such *Equal contribution. †Corresponding author. 1Code is available at: https://github.com/wangck20/DFML . Ensemble - based DML Backbone DFML …… Diverse Objectives Factorized MetricsFigure 1. Comparisons between ensemble-based deep metric learning methods and DFML. Ensemble-based DML learns an ensemble of embeddings where diverse objectives are employed. Differently, DFML factorizes the backbone and learns a certain routine for each sample to achieve the diversity of features, which further boosts the generalization ability of the model on unseen classes. (Best viewed in color.) as semantic instance segmentation [7, 21, 37], remote sens- ing [5, 10, 31], and room layout estimation [77]. Modern metric learning methods [44, 55, 56, 78] usually exploit deep neural networks to map an image to a single embedding and use the Euclidean distance or cosine simi- larity between embeddings to measure the similarity. As a single embedding might not be able to fully characterize an image, a number of methods [1, 43, 47, 49, 72, 79, 80] begin to explore using an ensemble of embeddings to represent an image, where each embedding describes one attribute of the image. The key to ensemble-based methods lies in how to enforce diversity in the ensemble of embeddings so that they can capture more characteristics. They achieve this by using a diversity loss [47, 49], selecting different samples [53, 72, 80], and adopting various tasks [43, 79], etc. Most existing methods adopt a shared backbone net- work to extract a common feature and only apply a single fully connected layer to obtain each specialized embedding. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7672 However, the shared backbone limits the diversity of the en- semble and hinders its ability to capture more generalizable features. It still receives a mix of all the training signals and can hardly produce diverse embeddings. To address this, we propose a deep factorized metric learning (DFML) method to adaptively factorize the train- ing signals to learn more generalizable features, as shown in 1. We first factorize each block of the metric back- bone model to a number of sub-blocks, where we make the summed features of all the sub-blocks to be equal to that of the full block. As different samples may possess distinct characteristics [80], we devise a learnable router to adaptively allocate the training samples to the correspond- ing sub-blocks. We learn the router using a reconstruction objective to encourage each sample to be processed by the most consistent sub-block. We demonstrate the proposed DFML framework is compatible with existing deep met- ric learning methods with various loss functions and sam- pling strategies and can be readily applied to them. Due to the better modularity of vision transformers (ViTs) [15,61], we mainly focus on factorizing ViTs and further bench- mark various existing deep metric learning methods on ViTs. Extensive experiments on the widely used CUB-200- 2011 [63], Cars196 [35], and Stanford Online Products [56] datasets show consistent improvements of DFML over ex- isting methods. We also provide an in-depth analysis of the proposed DFML framework to verify its effectiveness. Specifically, we show that backbone models trained by our DFML achieve better accuracy/computation trade-off than the original model on ImageNet-1K [52] and even improve the performance of ViT-B (+0.2% accuracy) with less com- putation load (-24% FLOPs).
Wang_FrustumFormer_Adaptive_Instance-Aware_Resampling_for_Multi-View_3D_Detection_CVPR_2023	Abstract The transformation of features from 2D perspective space to 3D space is essential to multi-view 3D object de- tection. Recent approaches mainly focus on the design of view transformation, either pixel-wisely lifting perspective view features into 3D space with estimated depth or grid- wisely constructing BEV features via 3D projection, treat- ing all pixels or grids equally. However, choosing what to transform is also important but has rarely been discussed before. The pixels of a moving car are more informative than the pixels of the sky. To fully utilize the informa- tion contained in images, the view transformation should be able to adapt to different image regions according to their contents. In this paper, we propose a novel framework named FrustumFormer , which pays more attention to the features in instance regions via adaptive instance-aware re- sampling. Specifically, the model obtains instance frustums on the bird’s eye view by leveraging image view object pro- posals. An adaptive occupancy mask within the instance frustum is learned to refine the instance location. More- over, the temporal frustum intersection could further reduce the localization uncertainty of objects. Comprehensive ex- periments on the nuScenes dataset demonstrate the effec- tiveness of FrustumFormer, and we achieve a new state-of- the-art performance on the benchmark. Codes and mod- els will be made available at https://github.com/ Robertwyq/Frustum .	1. Introduction Perception in 3D space has gained increasing attention in both academia and industry. Despite the success of LiDAR- based methods [14, 33, 41, 44], camera-based 3D object de- tection [19, 35, 36, 43] has earned a wide audience, due to the low cost for deployment and advantages for long-rangedetection. Recently, multi-view 3D detection in Bird’s-Eye- View (BEV) has made fast progresses. Due to the unified representation in 3D space, multi-view features and tem- poral information can be fused conveniently, which leads to significant performance improvement over monocular methods [5, 28, 35, 39]. Transforming perspective view features into the bird’s- eye view is the key to the success of modern BEV 3D de- tectors [12,18,19,22]. As shown in Fig. 1, we categorize the existing methods into lifting-based ones like LSS [30] and BEVDet [12] and query-based ones like BEVFormer [19] and Ego3RT [25]. However, these methods mainly focus on the design of view transformation strategies while over- looking the significance of choosing the right features to transform during view transformation. Regions containing objects like vehicles and pedestrians are apparently more in- formative than the empty background like sky and ground. But all previous methods treat them with equal importance. We suggest that the view transformation should be adaptive with respect to the image content. Therefore, we propose Adaptive Instance-aware Resampling (AIR) , an instance- aware view transformation, as shown in Fig. 1c. The core idea of AIR is to reduce instance localization uncertainty by focusing on a selective part of BEV queries. Localizing in- stance regions is difficult directly on the BEV plane but rel- atively easy in the image view. Therefore, the instance frus- tum, lifting from instance proposals in image views, gives geometrical hints of the possible locations of objects in the 3D space. Though the instance frustum has provided initial prior locations, it is still a large uncertain area. We propose anoccupancy mask predictor and a temporal frustum fusion module to further reduce the localization uncertainty. Our model learns an occupancy mask for frustum queries on the BEV plane, predicting the possibility that a region might contain objects. We also fuse instance frustums across dif- ferent time steps, where the intersection area poses geomet- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5096 (a) Grid Sampling in Image. (b) Grid Sampling in BEV . (c) Instance-aware Sampling in Frustum. Figure 1. Comparison of different sampling strategies for the feature transformation from image view to bird’s eye view. (a) represents the sampling in image view and lift features [12] to BEV plane with pixel-wise depth estimation. (b) shows the grid sampling in BEV and queries back [19] to obtain image features via cross-attention. (c) illustrates our proposed instance-aware sampling strategy in the frustum, which adapts to the view content by focusing more attention on instance regions. This approach is designed to enhance the learning of instance-aware BEV features. ric constraints for actual locations of objects. We propose a novel framework called FrustumFormer based on the insights mentioned previously, which effec- tively enhances the learning of instance-aware BEV features via Adaptive Instance-aware Resampling. FrustumFormer utilizes the instance frustum to establish the connection be- tween perspective and bird’s eye view regions, which con- tains two key designs: (1) A frustum encoder that enhances instance-aware features via adaptive instance-aware resam- pling. (2) A temporal frustum fusion module that aggre- gates historical instance frustum features for accurate local- ization and velocity prediction. In conclusion, the contribu- tions of this work are as follows: • We propose FrustumFormer , a novel framework that exploits the geometric constraints behind perspective view and birds’ eye view by instance frustum. • We propose that choosing what to transform is also im- portant during view transformation. The view transfor- mation should adapt to the view content. Instance re- gions should gain more attention rather than be treated equally. Therefore, we design Adaptive Instance- aware Resampling (AIR) to focus more on the instance regions, leveraging sparse instance queries to enhance the learning of instance-aware BEV features. • We evaluate the proposed FrustumFormer on the nuScenes dataset. We achieve improved performance compared to prior arts. FrustumFormer achieves 58.9 NDS and 51.6 mAP on nuScenes test set without bells and whistles.
Wang_Generalized_UAV_Object_Detection_via_Frequency_Domain_Disentanglement_CVPR_2023	Abstract When deploying the Unmanned Aerial Vehicles object detection (UAV-OD) network to complex and unseen real- world scenarios, the generalization ability is usually re- duced due to the domain shift. To address this issue, this paper proposes a novel frequency domain disentanglement method to improve the UAV-OD generalization. Specifi- cally, we first verified that the spectrum of different bands in the image has different effects to the UAV-OD general- ization. Based on this conclusion, we design two learnable filters to extract domain-invariant spectrum and domain- specific spectrum, respectively. The former can be used to train the UAV-OD network and improve its capacity for generalization. In addition, we design a new instance-level contrastive loss to guide the network training. This loss enables the network to concentrate on extracting domain- invariant spectrum and domain-specific spectrum, so as to achieve better disentangling results. Experimental re- sults on three unseen target domains demonstrate that our method has better generalization ability than both the base- line method and state-of-the-art methods.	1. Introduction Unmanned Aerial Vehicles (UA V) equipped with cam- eras have been exploited in a wide variety of applications, opening up a new frontier for computer vision applications [7, 11, 22, 28]. As one of the fundamental functions for the UA V-based applications, UA V object detection (UA V-OD) has garnered considerable interest [23, 31, 38]. However, the large mobility of UA V-mounted cameras leads to an un- predictable operating environment. The domain shift that occurs when applying a UA V-OD network that has been trained on a given dataset ( i.e., source domain) to unseen real-world data ( i.e., target domain) typically results in in- ∗Corresponding author. This work was supported by the National Key R&D Program of China under Grant 2020AAA0105702, the National Natural Science Foundation of China (NSFC) under Grants 62225207, 62276243 and U19B2038, the University Synergy Innovation Program of Anhui Province under Grant GXXT-2019-025. (a) Baseline (b) Ours Figure 1. Detection results on unseen target domains. UA V-OD network is trained on daylight images and tested on images with various scene structures (1st row), diverse illumination conditions (2nd row), and adverse weather conditions (3rd row). Green rect- angular boxes denote new correct detections beyond the baseline. adequate performance. In particular, unseen real-world data consists of unexpected and unknown samples, such as im- ages taken in various scene structures, diverse illumination conditions, and adverse weather conditions. Therefore, it is crucial to improve the generalization ability of UA V-OD. To alleviate the domain shift impact, existing methods broadly come in two flavors: Domain Adaptation (DA) [3, 5, 8, 16, 17, 26, 37] and Domain Generalization (DG) [19,20,27,30,40]. In general, DA aims to tackle the domain shift problem by learning domain-invariant/aligned features between the source and target domains. However, DA meth- ods cannot be readily employed when it is hard to guarantee the accessibility of the target data. The requirement to ac- cess both source and target data restricts the applicability of DA approaches. Recently, considerable attention has been drawn to the field of DG. The goal of DG is to learn a model using data from a single or multiple related but distinct source domains so that the model can generalize well under distri- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1064 Reject bandVarious Scene Diverse Illumination Adverse Weather Average AP50 AP75 AP AP50 AP75 AP AP50 AP75 AP AP50 AP75 AP Null (full band) 66.0 37.6 36.7 11.1 3.4 4.8 42.3 14.9 19.6 39.8 18.6 20.4 α= 0, β= 0.01 60.0 30.2 32.6 6.4 1.9 2.75 39.5 16.1 19.0 35.3 16.1 18.1 α= 0.01, β= 0.1 61.4 30.3 32.8 39.1 15.9 19.8 42.6 18.6 20.8 47.7 21.6 24.5 α= 0.1, β= 1 70.2 35.1 37.1 29.4 10.2 13.6 38.2 10.6 16.7 45.9 18.6 22.5 Table 1. We conduct preliminary experiments to explore whether different spectral bands contribute equally to the UA V-OD network’s generalization ability. The specified bands of source domain images are filtered out for training according to the reject band. For testing, the generalization performance of the UA V-OD network is evaluated on three unseen target domains. We adopt the evaluation protocols AP50, AP 75, and AP. ”Average” refers to the average generalization performance across three unseen target domains. We can conclude that eliminating various bands has distinct effects on the generalization of unseen target domains for UA V-OD network. bution shifts [43]. Most existing DG methods [19, 30, 40] focus on decoupling object-related features from global features via spatial vanilla convolution. However, unlike generic object detection scenarios based on surveillance or other ground-based cameras, the rapid movement of UA V- mounted cameras leads to severe changes in the global ap- pearance. For UA V-OD scenarios where the global appear- ance changes, it is essential to explore global dependency for better disentanglement. The spatial vanilla convolution, which only emphasizes local pixel attention, cannot fully explore global dependency, leading to suboptimal disentan- glement and generalization results. Inspired by the spectral theorem that the frequency do- main obeys the nature of global modeling, we propose to improve the UA V-OD generalization ability via frequency domain disentanglement. We first conduct preliminary ex- periments, i.e., exploring whether all spectrum bands con- tribute equally to the generalization for the UA V-OD task, to gain insight into how to implement our idea. If not, we can extract the spectrum that is conducive to generalization and use it to train the UA V-OD network to enhance its general- ization. Specifically, we first convert each source domain image x∈RH×W×Cinto frequency space through Fast Fourier Transform (FFT) [24]: F(x)(u, v) =H−1X h=0W−1X w=0x(h, w)e−j2π(h Hu+w Wv).(1) The frequency space signal F(x)can be further decom- posed to an amplitude spectrum A(x)and a phase spectrum P(x), which is expressed as: A(x)(u, v) = R2(x)(u, v) +I2(x)(u, v)1/2, P(x)(u, v) = arctanI(x)(u, v) R(x)(u, v) ,(2) where R(x)andI(x)represent the real and imaginary part ofF(x). For each source image, we filter out the bands of the amplitude spectrum A(x)between a certain upper threshold αand lower threshold β(’Reject band’ in Tab. 1)with a band reject filter fs∈RH×W×Cand obtain the re- maining amplitude spectrum ˆA(x): fs(i, j) =  1,i∈[αH 2,βH 2]∪[(1−α)H 2,(1−β)H 2] j∈[αW 2,βW 2]∪[(1−α)W 2,(1−β)W 2] 0,otherwise(3) A(x) = ˆA(x)⊗fs, (4) where ⊗denotes element-wise multiplication. ˆA(x)is then fed to Inverse Fast Fourier Transform (IFFT) with P(x)to generate the remaining image ˆxwhich is utilized to train the UA V-OD network. After training, we apply the UA V-OD network to three unseen target domains to evaluate the gen- eralization ability. The experimental results are presented in Tab. 1. We can observe that removing different bands has varying effects on generalization to three unseen target domains. Therefore, we can conclude that different bands contribute differently to the UA V-OD generalization. Based on the above observation, we employ two learn- able filters to identify and extract the domain-invariant and domain-specific spectrums. The former contributes posi- tively to generalization, while the latter is the opposite. Fur- thermore, we design a new instance-level contrastive loss to aid in learning the learnable filters, enabling them to concentrate on disentangling the two different spectrums. By optimizing the instance-level contrastive loss, the in- stance features of those are encouraged to contain domain- invariant characteristics shared by target objects, and the domain-specific characteristics shared in the source do- main, respectively. In this way, the UA V-OD network can generalize well on unseen target domains. For experiment settings, we focus on learning a single-domain generalized UA V-OD network, which is more challenging [30]. We fur- ther validate the network on three unseen target domains, in- cluding various scene structures, diverse illumination con- ditions, and adverse weather conditions, demonstrating su- perior generalization ability, as shown in Fig. 1. Our main contributions are highlighted as follows: 1065 • We provide a new perspective to improve the general- ization ability of the UA V-OD network on unseen tar- get domains. To our best knowledge, this is the first attempt to learn generalized UA V-OD via frequency domain disentanglement. • Based on the frequency domain disentanglement, we propose a new framework that utilizes two learnable filters to extract the domain-invariant and domain- specific spectrum and design an instance-level con- trastive loss to guide the disentangling process. • Extensive experiments on three unseen target domains reveal that our method enables the UA V-OD network to achieve superior generalization performance in com- parison to the baseline and state-of-the-art methods.